Current AI rankings

Best AI Models & Tools

Find the best AI for what you want to create.

The one to beat

Best Coding Overall

Winner · #1

Claude Code + Opus 5.5 High

Anthropic / Claude Code

Demanding implementation and everyday codebase changes

Model API$4 in $20 out per 1M tokens

* Early evidence — position may changeTested configuration

Why it wins & tradeoffs

An early lead at High effort in a six-project comparison, backed by benchmarks and daily-use reports. Serious failures keep the recommendation provisional.

Price & access: Claude paid plan or API billing; select Opus 5.5 High. Cost per finished task varies by workload.
Watch out for: Keep human review and tests in the loop. Reports include broken workflows; long-running unattended reliability is not established.
Content restrictions: Moderate
Evidence: A six-project test and Artificial Analysis favor Opus over Astra at High effort. Daily-use reports include both useful patches and serious failures; broader replication is missing.
System version: September 2026 Claude Code; exact build undisclosed in external test
Tested access: Claude paid plan or API billing; High effort; default safeguards

Sources and observations (4)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code with Opus 5.5 High leads the six-project aggregate, ahead of Codex Astra High, with lower measured average API cost and task time. Perfect published behavioral subtotal; strong quality grading. Actual evaluation dated September 23. Limited projects and repetitions, automated quality judging, harness build undisclosed; publisher-specific sponsorship unknown. This directly tests High in Claude Code; not a debugging-specialty result.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26 comparison: Opus 5.5 adaptive High with default fallback leads Astra High on Terminal-Bench 4.0 and SciCode, but Astra xhigh remains ahead of Opus High on Terminal-Bench. AA's model evaluation is contextual for Claude Code, not a Claude Code harness score. Intelligence-index cost per task is slightly higher than Astra High despite lower token prices; do not promise universally cheaper completion.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Opus 5.5 High: creator preferred its outputs on all three builds for detail, prompt fidelity and gameplay. Supports creative coding quality; does not establish superiority on repository repair or every coding task. Creator still prefers Astra for everyday refactors/architecture. Scope: creative coding support within the broader coding rubric; repository performance remains unmeasured here.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 Vibe Check: original Claude Code/app tests support daily implementation and concise patches, with some builders switching from Fable/Codex. Contrary evidence includes broken core app flows, uncalled required service and destructive verification behavior. Examples use mixed efforts (including xhigh), so they are contextual for High, not matched High victories. Every received Anthropic pre-launch access; editorial independence asserted by Every, no independent-replication claim.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

New evidence since this ranking (1)

These findings have not yet changed the ranking. Several articles from one publisher count as one source.

AI Coding Daily — original coding benchmarkSource record · Reviewed Oct 3, 2026
October 3 table: Opus 5.5 High / Claude Code; $0.79, 3m10s, 67.41/70. Added bug-finding task changes denominator from 60; not a longitudinal improvement. Exact build and added-task rubric incompletely disclosed. One publisher; no standalone debugging qualification. Preserve prior observations.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 23, 2026Provider, pricing & access ↗

Codex + GPT-6 Astra

OpenAI / Codex

Codebase and terminal work with adjustable reasoning effort

* Early evidence — position may changeTested configuration

When to choose Codex + GPT-6 Astra

A close alternative that completed the test's required behaviors but trailed Opus at High effort. Extra-high terminal results are a separate strength.

Model API$10 in $50 out per 1M tokens

Price & access: Codex, ChatGPT Work or API; access and cost depend on plan and effort.
Watch out for: Opus leads the reviewed six-project test at High effort. Astra's stronger terminal results use a different effort setting.
Content restrictions: Low
Evidence: Opus leads the reviewed High-versus-High coding comparisons. Astra leads separate terminal tests at extra-high effort; those settings are not interchangeable.
System version: September 2026 production harness
Tested access: Codex, ChatGPT Work, or API

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot explicitly tests Codex CLI + GPT-6 Astra High and Medium separately. High trails Claude Code + Opus 5.5 High on aggregate quality; both clear the behavior tasks. High is slower and costlier on this suite. High evidence is a configuration-specific observation within the historically broad Astra entry, not evidence about xhigh/max settings.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26: Astra High trails Opus 5.5 High on terminal and scientific coding, while Astra xhigh retains a terminal advantage over Opus High. AA evaluates a different harness from Codex; settings must stay explicit. Astra retains stronger AutomationBench results. The broad Astra configuration supports choosing xhigh for hard terminal work, not claiming its High setting wins.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Codex + GPT-6 Astra, explicitly High in this comparison: Opus won creator preference on all three creative builds; Astra was faster in reported runs and cheapest on two. Creator still favors Astra for everyday refactors/architecture, a reported workflow preference rather than a measured debugging test. Do not transfer to xhigh or other efforts. Scope: creative coding support within the broader coding rubric; repository performance remains unmeasured here.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review retains Astra for time-bounded work and reports higher raw throughput in one coding experiment, while Opus met more latency budgets. Scope is one experiment with mixed settings, not an Astra High benchmark. September 3 production-rewrite benchmark remains separate and used Astra extra-high.
Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
Both configurations completed substantial practical work. The comparison favored Codex + Astra more often for day-to-day execution while preserving Claude Code + Fable 5.1 as stronger on selected tasks.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Claude Code + Fable 5.1

Anthropic / Claude Code

Difficult implementation and careful changes to existing code

* Early evidence — position may changeTested configuration

When to choose Claude Code + Fable 5.1

A strong choice for difficult codebase problems, though newer comparisons favor other models. Tests use different effort settings.

Model API$10 in $50 out per 1M tokens

Price & access: Claude subscription or API billing; heavier usage can be costly.
Watch out for: Newer models lead some implementation tests and may cost less per task.
Content restrictions: Moderate
Evidence: Every still prefers Fable for some hard tasks. Other comparisons use Medium or different effort settings, so they do not establish a uniform defeat.
System version: September 2026 production release
Tested access: Claude Max or API billing

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code + Fable 5.1 Medium trails Opus 5.5 High and Astra High on the six-project aggregate. The Fable row is Medium: it does not measure Fable High/Max. Retain that scope and use separate evidence for hard-problem frontier capability.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 22 original Opus 5.5 analysis: Opus max surpasses Fable 5.1's prior SciCode result and overall index, but this max comparison is contextual and cannot be relabeled Opus High versus Fable High. Fable remains a frontier coding model with premium pricing. Retain the older direct Fable configuration observations alongside this version-specific challenge.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Fable 5.1, explicitly High in this comparison: Opus won creator preference across all three creative builds. Fable showed speed but missed some requirements. These are task-level observations, not a universal third-place claim or evidence for other effort settings. Scope: creative coding support within the broader coding rubric; repository performance remains unmeasured here.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review: some daily coding moves to Opus for value, but two testers still prefer Fable 5.1 on the hardest coding problems. Ruby comparison used Fable High versus Opus xhigh; do not treat effort as matched. This supports retaining Fable as a demanding-task alternative. Anthropic supplied pre-launch access to Opus; keep that relationship disclosed.
Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
Both configurations completed substantial practical work. The comparison favored Codex + Astra more often for day-to-day execution while preserving Claude Code + Fable 5.1 as stronger on selected tasks.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

What we judge

Our strongest all-around choice for planning, building, testing, and maintaining code. We weigh quality, reliability, workflow, and cost together.

CorrectnessRepository understandingTool useMaintainabilityWorkflow reliability

About this provisional ranking

Based on external tests and hands-on comparisons. Results depend on the model, reasoning effort and coding tools. More testing may change the order; reviews are due every two days.

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (57)

The order is unchanged; some comparisons are still missing.

Terra effort should follow the task. Bug Hunt Bench's October 26–27 runs found 40.7 of 105 planted bugs at Max versus 11.3 at XHigh, averaged over three attempts; NerdsChalk's much smaller Python repair passed at every tested effort from Low through XHigh. These distinct tasks favor testing Max on difficult repair rather than paying for it automatically.

Next comparison: Compare exact Terra CLI 0.159.0 efforts with current leaders on production-scale repairs, tool failures and actual cost; seek independent production Hy4/OpenCode evidence.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Current High-effort implementation remains competitive, but the older Astra agentic/refactoring rationale mixes effort settings and dates. Keep it as a historical recommendation, not proof that every Codex Astra setting defeats newer contenders.
Claude Code + Fable 5.1
Keep Fable 5.1's historical careful-codebase rationale with its original dates. The new effort table uses Medium, so it cannot settle High/Max performance or refresh the old general configuration.
Claude Code + Opus 5
Opus 5 remains the historical baseline; this pass adds no current exact-version advantage over Opus 5.5 or Sol 6.1. Retention is not a new win.
Cursor + Fable 5.1
Cursor + Fable 5.1 needs a Cursor-specific result; Claude Code findings cannot establish this configuration's quality or agent reliability.
Codex + GPT-5.6 Sol
GPT-5.6 Sol is an older Codex baseline. New Sol 6.1 findings are not transferable evidence for this version.
Replit Agent + Automatically Routed
Automatic routing offers an integrated workflow, but the routed stack and cost per completed task must be disclosed before model-specific quality attribution.
GitHub Copilot + Automatically Routed
Copilot Auto's convenient workflow does not identify which model completed each task. Require current routing, intervention and task-completion records.
Devin + SWE-2
Devin + SWE-2 has saved mixed repair/terminal results; the older vendor and selected-suite evidence cannot establish a current across-task lead.
Claude Code + Opus 5.5 High
Opus 5.5 High remains a reasonable implementation shortlist choice: recent short-task and practical-build findings support it, while intervention and sustained-task reliability remain limitations.
Codex + GPT-6 Sol High
GPT-6 Sol High is a separate predecessor to Sol 6.1. Its saved evidence is retained without borrowing successor results.
Claude Code + Sonnet 5.5 High
Sonnet 5.5 High is represented in the current short-task comparison. A credible lower-cost candidate, but not a demonstrated all-workload replacement for the frontier choices.
Codex + GPT-6.1 Sol High
Sol 6.1 High is a credible quality/value shortlist alternative. Current short-task evidence is close to the leaders, with narrow reconstruction support; exact independent replication and broad reliability remain open.
Codex + GPT-6.1 Sol Medium
Sol 6.1 Medium is a sensible speed/value setting to consider for bounded changes. Its weaker-effort result is its own tradeoff, not proof of High-level quality.
GitHub Copilot + GPT-6.1 Sol
Copilot's Sol 6.1 default setting needs Copilot-specific observation. Codex effort results cannot establish its behavior.
GitHub Copilot HydraFusion
September HydraFusion cannot inherit results for the October adaptive or Max releases; record the routing and build used by any comparison.
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)
Replit Max Mode changes harness and effort selection. Codex Sol High findings do not establish this usable configuration's quality or cost.
Sonnet 5.5 + Replit Agent Power (recommended effort)
Replit Power with Sonnet requires its own completion and recovery evidence; Claude Code Sonnet High is a different configuration.
Step 5 Preview
Step 5 Preview's coding eligibility does not establish a reproducible agent/harness configuration or comparative repository performance.
Claude Code + Claude Haiku 5.5 High
Haiku 5.5 High is a separate effort configuration requiring task-specific correctness and recovery evidence; speed or vendor availability alone is insufficient.
Claude Code + Claude Haiku 5.5 Medium
Haiku 5.5 Medium needs its own observed tradeoff against High; neither effort inherits the other's result.
Codex + GPT-6.1 Sol Low
Sol 6.1 Low is worth considering for quick, bounded changes. Initial effort findings suggest weaker quality than High; require review and tests rather than treating lower cost as equal capability.
Cursor + Grok 4.7 High
Cursor + Grok 4.7 High is a usable challenger. Available narrow benchmark signals require artifact, funding and workload checks before a broad editorial order.
Mellum2.1 Thinking · vLLM · documented thinking settings
Mellum2.1 Thinking in vLLM needs actual repository tasks with the documented thinking parameters, not model-level or completion-only evidence.
Muse Wonder · Muse Code · provider defaults
Muse Wonder defaults are distinct from Muse Wonder Full High goal mode. The latter's practical failures are not evidence that this exact configuration fails.
Muse Wonder Full · Muse Code · provider defaults
Muse Wonder Full defaults cannot inherit the High goal-mode result. Default execution, quota behavior and task recovery still require direct observation.
Gemini 4 Argon · Antigravity September 2026 · Medium
Argon in Antigravity Medium needs harness-matched repository outcomes; Cursor, Copilot and bare-model results are not interchangeable.
Gemini 4 Argon · GitHub Copilot CLI · provider default
Argon in Copilot's provider default needs recorded default/routing behavior and actual task outcomes.
Muse Wonder Full · Muse Code · High goal mode
Pat tested High goal mode with Contributor off; VulcanBench High used Contributor on in 1.4.9. Preserve Pat’s interrupted quota and estimated cost. These are not a replication.
Gemini 4 Argon · Cursor · High
Argon Cursor High requires its own exact-effort outcomes and cost/intervention record.
Gemini 4 Argon · Cursor · Medium
Argon Cursor Medium must be assessed independently from High; no effort uplift is inferred.
GitHub Copilot HydraFusion · October 14
October 14 HydraFusion adaptive routing is a materially different stack from September. Require disclosed routed models and repeatable task records.
GitHub Copilot HydraFusion Max · October 14
HydraFusion Max needs evidence for its own maximum-quality routing and usage costs, not automatic inheritance from adaptive mode.
OpenCode + Beam 1.0 · Max
Beam 1.0 OpenCode Max is in the usable field. Current short-task signals do not resolve independent exact-configuration reliability and completion quality.
OpenCode + Nemotron 3.5 Super VL · High
Nemotron 3.5 Super VL OpenCode High needs corroborated repository outcomes with exact routing/build details.
OpenCode + Mistral Large 4 · High
Mistral Large 4 OpenCode High needs task-matched independent replication before a broad placement.
DeepSeek Harness + V4.1 Flash 1017 · Max
DeepSeek V4.1 Flash 1017 in the native harness is distinct from OpenCode. Assess each one's completion, failures and costs separately.
OpenCode + DeepSeek V4.1 Flash 1017 · Max
DeepSeek Flash 1017 OpenCode Max is a separate comparison arm from native-harness Max; shared model identity is not shared agent performance.
OpenCode + DeepSeek V4.1 Flash 1017 · High
DeepSeek Flash 1017 OpenCode High cannot inherit Max's measured quality or resource use.
Claude Code + MiniMax M3.1 Flash (Max, native MiniMax)
Native MiniMax M3.1 Flash Max through Claude Code needs route-specific evidence; third-party or OpenCode runs may behave differently.
OpenCode + MiniMax M3.1 Flash · Max
MiniMax M3.1 Flash OpenCode Max needs disclosed route, exact effort and reproducible outcomes before comparative qualification.
OpenCode + Recursive Flint · rec 4-bit · Default
Recursive Flint rec 4-bit is a precision-specific configuration. Results from a different quantization or provider route cannot establish its coding quality.
Kimi K3.1 Pro + Kimi Code — default effort
Kimi K3.1 Pro Kimi Code default is not the same arm as OpenCode Max or explicit Kimi Code High. Resolve the actual default before comparing.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low
Atomic Agent 0.6.9 Kimi Low needs its own task record; Kimi Code and OpenCode results do not transfer.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max
Atomic Agent 0.6.9 Kimi Max must be compared separately from Low and from other Kimi harnesses.
Codex + GPT-6.1 Terra — provider default
Terra's fresh benchmark page reports explicit effort arms; these cannot rank the registered provider-default Codex configuration. Determine its actual effort and match the usable configuration first.
GitHub Copilot Fusion · Copilot CLI · Standard
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · High
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · Max
The Oct 28 original AA test verified Fusion Max in the Copilot CLI: led terminal tasks, trailed Opus Max overall and on repository questions with a lower estimated cost per task. Organization-specific routing, intervention cost and independent same-configuration work remain open.
Particle-1 · Codex via Vercel AI Gateway · provider default
Gateway access is verified. No matched repository comparison establishes whether token savings survive retries or preserve correctness.
Codex + GPT-6.1 Terra — XHigh
One partially reviewed October 27 BuilderBench run used Terra XHigh in Codex. The comparison used a different harness for Kimi; neither costs nor general performance are settled.
Muse Wonder Full · Muse Code Contributor · High
VulcanBench's Oct 13–17 reconstruction suite tested High in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Extra-high
VulcanBench's Oct 13–17 reconstruction suite tested Extra-high in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Low
VulcanBench's Oct 13–17 reconstruction suite tested Low in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. 18 of 23 tasks fully solved. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4
Resolved 40 of 50 offline SWE-bench Verified instances on one RTX PRO 6000. Same seeded subset has not been run on current podium rivals, so not a head-to-head win.
GPT-6.1 Terra · Codex CLI 0.159.0 · Low
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Medium
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · High
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Max
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
OpenCode + Tencent Hy4 · OpenRouter default
OpenCode can route to production Tencent Hy4. No matched repository run verified at default effort.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Claude Code + Claude Haiku 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Claude Haiku 5.5 MediumUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + MiniMax M3.1 Flash (Max, native MiniMax)Unresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Sonnet 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol LowUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol MediumUnresolved since Oct 22, 2026 · Priority reassessment
Cursor + Grok 4.7 HighUnresolved since Oct 22, 2026 · Priority reassessment
DeepSeek Harness + V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Antigravity September 2026 · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · HighUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · GitHub Copilot CLI · provider defaultUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot + GPT-6.1 SolUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusionUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion Max · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Mellum2.1 Thinking · vLLM · documented thinking settingsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · High goal modeUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Beam 1.0 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + MiniMax M3.1 Flash · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Mistral Large 4 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Nemotron 3.5 Super VL · HighUnresolved since Oct 22, 2026 · Priority reassessment
Sonnet 5.5 + Replit Agent Power (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Step 5 PreviewUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Recursive Flint · rec 4-bit · DefaultUnresolved since Oct 22, 2026 · Priority reassessment
Kimi K3.1 Pro + Kimi Code — default effortUnresolved since Oct 23, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — LowUnresolved since Oct 25, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — MaxUnresolved since Oct 25, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — provider defaultUnresolved since Oct 26, 2026 · Priority reassessment
Claude Code + Opus 5Unresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-5.6 SolUnresolved since Oct 29, 2026 · Priority reassessment
Cursor + Fable 5.1Unresolved since Oct 29, 2026 · Priority reassessment
Devin + SWE-2Unresolved since Oct 29, 2026 · Priority reassessment
GitHub Copilot + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — XHighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · HighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · MaxUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · StandardUnresolved since Oct 31, 2026
Particle-1 · Codex via Vercel AI Gateway · provider defaultUnresolved since Oct 31, 2026
Muse Wonder Full · Muse Code Contributor · Extra-highUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · HighUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · LowUnresolved since Nov 1, 2026
Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4Unresolved since Nov 2, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · HighUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · LowUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MaxUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MediumUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · XHighUnresolved since Nov 4, 2026
OpenCode + Tencent Hy4 · OpenRouter defaultUnresolved since Nov 4, 2026

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

51 eligible contenders are not covered by the published assessment. Their absence is not a quality verdict.

Review dates and coverage

Assessment published: Sep 26, 2026
Ranking order changed: Sep 26, 2026
Supporting observations: Sep 12, 2026 – Sep 26, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 9 of 60 eligible models

Still needing comparison: Codex + GPT-6 Sol High; Claude Code + Sonnet 5.5 High; Codex + GPT-6.1 Sol High; Codex + GPT-6.1 Sol Medium; GitHub Copilot + GPT-6.1 Sol; GitHub Copilot HydraFusion; GPT-6.1 Sol + Replit Agent Max Mode (recommended effort); Sonnet 5.5 + Replit Agent Power (recommended effort); Step 5 Preview; Claude Code + Claude Haiku 5.5 High; Claude Code + Claude Haiku 5.5 Medium; Codex + GPT-6.1 Sol Low; Cursor + Grok 4.7 High; Mellum2.1 Thinking · vLLM · documented thinking settings; Muse Wonder · Muse Code · provider defaults; Muse Wonder Full · Muse Code · provider defaults; Gemini 4 Argon · Antigravity September 2026 · Medium; Gemini 4 Argon · GitHub Copilot CLI · provider default; Muse Wonder Full · Muse Code · High goal mode; Gemini 4 Argon · Cursor · High; Gemini 4 Argon · Cursor · Medium; GitHub Copilot HydraFusion · October 14; GitHub Copilot HydraFusion Max · October 14; OpenCode + Beam 1.0 · Max; OpenCode + Nemotron 3.5 Super VL · High; OpenCode + Mistral Large 4 · High; DeepSeek Harness + V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · High; Claude Code + MiniMax M3.1 Flash (Max, native MiniMax); OpenCode + MiniMax M3.1 Flash · Max; OpenCode + Recursive Flint · rec 4-bit · Default; Kimi K3.1 Pro + Kimi Code — default effort; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max; Codex + GPT-6.1 Terra — provider default; GitHub Copilot Fusion · Copilot CLI · Standard; GitHub Copilot Fusion · Copilot CLI · High; GitHub Copilot Fusion · Copilot CLI · Max; Particle-1 · Codex via Vercel AI Gateway · provider default; Codex + GPT-6.1 Terra — XHigh; Muse Wonder Full · Muse Code Contributor · High; Muse Wonder Full · Muse Code Contributor · Extra-high; Muse Wonder Full · Muse Code Contributor · Low; Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4; GPT-6.1 Terra · Codex CLI 0.159.0 · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Medium; GPT-6.1 Terra · Codex CLI 0.159.0 · High; GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh; GPT-6.1 Terra · Codex CLI 0.159.0 · Max; OpenCode + Tencent Hy4 · OpenRouter default.

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Models, coding tools, and builders: what’s the difference?

The model: Astra is the AI doing the reasoning. The tools it can use depend on where you access it.
The coding tool: Codex gives a model tools to work on project files, run tests, and fix problems. “Codex + Astra” names the combination being compared.
The builder: A platform such as Lovable brings building, databases, and publishing into one service. We judge the whole experience, including how easily you can maintain the result.

And ChatGPT Work? It’s a broader workspace for completing tasks; Codex provides the coding capabilities. Work and Astra aren’t competing models. The model, available tools, and setup together affect the result.

A tool can appear in more than one category. A website primarily presents information; a web app lets people do things with data. Some projects do both.

Advanced comparisonsAutonomous coding · Writing code · Fixing & improving code

Why doesn’t the overall winner win every category?

Overall balances several kinds of work, reliability, workflow, and cost. A specialist can produce a more refined implementation or a better fix while another tool offers the strongest all-around combination. Builder rankings also judge the finished product and publishing experience, so their winners can differ.

How we weigh the evidence →

Autonomous Coding

Winner · #1

Codex + GPT-6 Astra

OpenAI / Codex

Codebase and terminal work with adjustable reasoning effort

Model API$10 in $50 out per 1M tokens

* Early evidence — position may changeTested configuration

Why it wins & tradeoffs

Leads for broader terminal and production work, including extra-high effort. Opus wins the reviewed coding comparisons when both use High effort.

Price & access: Codex, ChatGPT Work or API; access and cost depend on plan and effort.
Watch out for: Opus leads the reviewed six-project test at High effort. Astra's stronger terminal results use a different effort setting.
Content restrictions: Low
Evidence: Opus leads the reviewed High-versus-High coding comparisons. Astra leads separate terminal tests at extra-high effort; those settings are not interchangeable.
System version: September 2026 production harness
Tested access: Codex, ChatGPT Work, or API

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot explicitly tests Codex CLI + GPT-6 Astra High and Medium separately. High trails Claude Code + Opus 5.5 High on aggregate quality; both clear the behavior tasks. High is slower and costlier on this suite. High evidence is a configuration-specific observation within the historically broad Astra entry, not evidence about xhigh/max settings.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26: Astra High trails Opus 5.5 High on terminal and scientific coding, while Astra xhigh retains a terminal advantage over Opus High. AA evaluates a different harness from Codex; settings must stay explicit. Astra retains stronger AutomationBench results. The broad Astra configuration supports choosing xhigh for hard terminal work, not claiming its High setting wins.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Codex + GPT-6 Astra, explicitly High in this comparison: Opus won creator preference on all three creative builds; Astra was faster in reported runs and cheapest on two. Creator still favors Astra for everyday refactors/architecture, a reported workflow preference rather than a measured debugging test. Do not transfer to xhigh or other efforts. Scope: extended autonomous tool-using builds; not controlled repository repair. Deployment reminder/intervention limits one-shot claims.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review retains Astra for time-bounded work and reports higher raw throughput in one coding experiment, while Opus met more latency budgets. Scope is one experiment with mixed settings, not an Astra High benchmark. September 3 production-rewrite benchmark remains separate and used Astra extra-high.
OpenAI Codex changelogSource page · Observed Sep 18, 2026
Codex CLI 0.155.0 added live reasoning-status updates, agent-task archiving and worktree ownership, Touch ID verification for local MCP requests, recoverable daemon updates, and authentication recovery; 0.155.1 corrected reasoning-summary defaults for providers that reject them. These are material harness workflow and reliability changes, not direct model-quality evidence.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Claude Code + Opus 5.5 High

Anthropic / Claude Code

Demanding implementation and everyday codebase changes

* Early evidence — position may changeTested configuration

When to choose Claude Code + Opus 5.5 High

Strong High-effort coding results, but broken workflows and a deleted verification file limit confidence in sustained unattended work.

Model API$4 in $20 out per 1M tokens

Price & access: Claude paid plan or API billing; select Opus 5.5 High. Cost per finished task varies by workload.
Watch out for: Keep human review and tests in the loop. Reports include broken workflows; long-running unattended reliability is not established.
Content restrictions: Moderate
Evidence: A six-project test and Artificial Analysis favor Opus over Astra at High effort. Daily-use reports include both useful patches and serious failures; broader replication is missing.
System version: September 2026 Claude Code; exact build undisclosed in external test
Tested access: Claude paid plan or API billing; High effort; default safeguards

Sources and observations (4)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code with Opus 5.5 High leads the six-project aggregate, ahead of Codex Astra High, with lower measured average API cost and task time. Perfect published behavioral subtotal; strong quality grading. Actual evaluation dated September 23. Limited projects and repetitions, automated quality judging, harness build undisclosed; publisher-specific sponsorship unknown. This directly tests High in Claude Code; not a debugging-specialty result.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26 comparison: Opus 5.5 adaptive High with default fallback leads Astra High on Terminal-Bench 4.0 and SciCode, but Astra xhigh remains ahead of Opus High on Terminal-Bench. AA's model evaluation is contextual for Claude Code, not a Claude Code harness score. Intelligence-index cost per task is slightly higher than Astra High despite lower token prices; do not promise universally cheaper completion.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Opus 5.5 High: creator preferred its outputs on all three builds for detail, prompt fidelity and gameplay. Supports creative coding quality; does not establish superiority on repository repair or every coding task. Creator still prefers Astra for everyday refactors/architecture. Scope: extended autonomous tool-using builds; not controlled repository repair. Deployment reminder/intervention limits one-shot claims.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 Vibe Check: original Claude Code/app tests support daily implementation and concise patches, with some builders switching from Fable/Codex. Contrary evidence includes broken core app flows, uncalled required service and destructive verification behavior. Examples use mixed efforts (including xhigh), so they are contextual for High, not matched High victories. Every received Anthropic pre-launch access; editorial independence asserted by Every, no independent-replication claim.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 23, 2026Provider, pricing & access ↗

Claude Code + Fable 5.1

Anthropic / Claude Code

Difficult implementation and careful changes to existing code

* Early evidence — position may changeTested configuration

When to choose Claude Code + Fable 5.1

A strong choice for difficult codebase problems, though newer comparisons favor other models. Tests use different effort settings.

Model API$10 in $50 out per 1M tokens

Price & access: Claude subscription or API billing; heavier usage can be costly.
Watch out for: Newer models lead some implementation tests and may cost less per task.
Content restrictions: Moderate
Evidence: Every still prefers Fable for some hard tasks. Other comparisons use Medium or different effort settings, so they do not establish a uniform defeat.
System version: September 2026 production release
Tested access: Claude Max or API billing

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code + Fable 5.1 Medium trails Opus 5.5 High and Astra High on the six-project aggregate. The Fable row is Medium: it does not measure Fable High/Max. Retain that scope and use separate evidence for hard-problem frontier capability.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 22 original Opus 5.5 analysis: Opus max surpasses Fable 5.1's prior SciCode result and overall index, but this max comparison is contextual and cannot be relabeled Opus High versus Fable High. Fable remains a frontier coding model with premium pricing. Retain the older direct Fable configuration observations alongside this version-specific challenge.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Fable 5.1, explicitly High in this comparison: Opus won creator preference across all three creative builds. Fable showed speed but missed some requirements. These are task-level observations, not a universal third-place claim or evidence for other effort settings. Scope: extended autonomous tool-using builds; not controlled repository repair. Deployment reminder/intervention limits one-shot claims.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review: some daily coding moves to Opus for value, but two testers still prefer Fable 5.1 on the hardest coding problems. Ruby comparison used Fable High versus Opus xhigh; do not treat effort as matched. This supports retaining Fable as a demanding-task alternative. Anthropic supplied pre-launch access to Opus; keep that relationship disclosed.
Claude Code changelogSource page · Observed Sep 18, 2026
Claude Code 2.1.274–2.1.276 added critical memory warnings, configurable MCP startup waits, synced claude.ai skills/plugins, marketplace plugin installation, improved queued-message control, and then repaired a proxy/gateway regression introduced in 2.1.275. These are material agent workflow and reliability changes, not direct task-quality benchmark evidence.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

What we judge

For delegating a multi-step development task: the agent plans, edits files, runs tests, and works through errors with limited supervision.

Task completionPlanningTool useReliability on long tasksRecovery from errors

About this provisional ranking

Based on external tests and hands-on comparisons. Results depend on the model, reasoning effort and coding tools. More testing may change the order; reviews are due every two days.

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (56)

The order is unchanged; some comparisons are still missing.

Terra Max's stronger reconstruction and planted-bug results support spending more effort on difficult repository work. Small repair tasks can saturate at lower effort. BuilderBench's original configured agent trials remain separate; neither a bare-model score nor an offline suite establishes broad autonomous reliability.

Next comparison: Match agent, CLI version and effort against current leaders on multi-step repository tasks, recovery and cost.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Current High-effort implementation remains competitive, but the older Astra agentic/refactoring rationale mixes effort settings and dates. Keep it as a historical recommendation, not proof that every Codex Astra setting defeats newer contenders. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Fable 5.1
Keep Fable 5.1's historical careful-codebase rationale with its original dates. The new effort table uses Medium, so it cannot settle High/Max performance or refresh the old general configuration. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Opus 5
Opus 5 remains the historical baseline; this pass adds no current exact-version advantage over Opus 5.5 or Sol 6.1. Retention is not a new win. Agentic placement additionally requires sustained completion and intervention evidence.
Cursor + Fable 5.1
Cursor + Fable 5.1 needs a Cursor-specific result; Claude Code findings cannot establish this configuration's quality or agent reliability. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-5.6 Sol
GPT-5.6 Sol is an older Codex baseline. New Sol 6.1 findings are not transferable evidence for this version. Agentic placement additionally requires sustained completion and intervention evidence.
Replit Agent + Automatically Routed
Automatic routing offers an integrated workflow, but the routed stack and cost per completed task must be disclosed before model-specific quality attribution. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot + Automatically Routed
Copilot Auto's convenient workflow does not identify which model completed each task. Require current routing, intervention and task-completion records. Agentic placement additionally requires sustained completion and intervention evidence.
Devin + SWE-2
Devin + SWE-2 has saved mixed repair/terminal results; the older vendor and selected-suite evidence cannot establish a current across-task lead. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Opus 5.5 High
Opus 5.5 High remains a reasonable implementation shortlist choice: recent short-task and practical-build findings support it, while intervention and sustained-task reliability remain limitations. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-6 Sol High
GPT-6 Sol High is a separate predecessor to Sol 6.1. Its saved evidence is retained without borrowing successor results. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Sonnet 5.5 High
Sonnet 5.5 High is represented in the current short-task comparison. A credible lower-cost candidate, but not a demonstrated all-workload replacement for the frontier choices. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-6.1 Sol High
Sol 6.1 High is a credible quality/value shortlist alternative. Current short-task evidence is close to the leaders, with narrow reconstruction support; exact independent replication and broad reliability remain open. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-6.1 Sol Medium
Sol 6.1 Medium is a sensible speed/value setting to consider for bounded changes. Its weaker-effort result is its own tradeoff, not proof of High-level quality. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot + GPT-6.1 Sol
Copilot's Sol 6.1 default setting needs Copilot-specific observation. Codex effort results cannot establish its behavior. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot HydraFusion
September HydraFusion cannot inherit results for the October adaptive or Max releases; record the routing and build used by any comparison. Agentic placement additionally requires sustained completion and intervention evidence.
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)
Replit Max Mode changes harness and effort selection. Codex Sol High findings do not establish this usable configuration's quality or cost. Agentic placement additionally requires sustained completion and intervention evidence.
Sonnet 5.5 + Replit Agent Power (recommended effort)
Replit Power with Sonnet requires its own completion and recovery evidence; Claude Code Sonnet High is a different configuration. Agentic placement additionally requires sustained completion and intervention evidence.
Step 5 Preview
Step 5 Preview's coding eligibility does not establish a reproducible agent/harness configuration or comparative repository performance. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Claude Haiku 5.5 High
Haiku 5.5 High is a separate effort configuration requiring task-specific correctness and recovery evidence; speed or vendor availability alone is insufficient. Agentic placement additionally requires sustained completion and intervention evidence.
Claude Code + Claude Haiku 5.5 Medium
Haiku 5.5 Medium needs its own observed tradeoff against High; neither effort inherits the other's result. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-6.1 Sol Low
Sol 6.1 Low is worth considering for quick, bounded changes. Initial effort findings suggest weaker quality than High; require review and tests rather than treating lower cost as equal capability. Agentic placement additionally requires sustained completion and intervention evidence.
Cursor + Grok 4.7 High
Cursor + Grok 4.7 High is a usable challenger. Available narrow benchmark signals require artifact, funding and workload checks before a broad editorial order. Agentic placement additionally requires sustained completion and intervention evidence.
Mellum2.1 Thinking · vLLM · documented thinking settings
Mellum2.1 Thinking in vLLM needs actual repository tasks with the documented thinking parameters, not model-level or completion-only evidence. Agentic placement additionally requires sustained completion and intervention evidence.
Muse Wonder · Muse Code · provider defaults
Muse Wonder defaults are distinct from Muse Wonder Full High goal mode. The latter's practical failures are not evidence that this exact configuration fails. Agentic placement additionally requires sustained completion and intervention evidence.
Muse Wonder Full · Muse Code · provider defaults
Muse Wonder Full defaults cannot inherit the High goal-mode result. Default execution, quota behavior and task recovery still require direct observation. Agentic placement additionally requires sustained completion and intervention evidence.
Gemini 4 Argon · Antigravity September 2026 · Medium
Argon in Antigravity Medium needs harness-matched repository outcomes; Cursor, Copilot and bare-model results are not interchangeable. Agentic placement additionally requires sustained completion and intervention evidence.
Gemini 4 Argon · GitHub Copilot CLI · provider default
Argon in Copilot's provider default needs recorded default/routing behavior and actual task outcomes. Agentic placement additionally requires sustained completion and intervention evidence.
Muse Wonder Full · Muse Code · High goal mode
Pat tested High goal mode with Contributor off; VulcanBench High used Contributor on in 1.4.9. Preserve Pat’s interrupted quota and estimated cost. These are not a replication.
Gemini 4 Argon · Cursor · High
Argon Cursor High requires its own exact-effort outcomes and cost/intervention record. Agentic placement additionally requires sustained completion and intervention evidence.
Gemini 4 Argon · Cursor · Medium
Argon Cursor Medium must be assessed independently from High; no effort uplift is inferred. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot HydraFusion · October 14
October 14 HydraFusion adaptive routing is a materially different stack from September. Require disclosed routed models and repeatable task records. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot HydraFusion Max · October 14
HydraFusion Max needs evidence for its own maximum-quality routing and usage costs, not automatic inheritance from adaptive mode. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + Beam 1.0 · Max
Beam 1.0 OpenCode Max is in the usable field. Current short-task signals do not resolve independent exact-configuration reliability and completion quality. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + Nemotron 3.5 Super VL · High
Nemotron 3.5 Super VL OpenCode High needs corroborated repository outcomes with exact routing/build details. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + Mistral Large 4 · High
Mistral Large 4 OpenCode High needs task-matched independent replication before a broad placement. Agentic placement additionally requires sustained completion and intervention evidence.
DeepSeek Harness + V4.1 Flash 1017 · Max
DeepSeek V4.1 Flash 1017 in the native harness is distinct from OpenCode. Assess each one's completion, failures and costs separately. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + DeepSeek V4.1 Flash 1017 · Max
DeepSeek Flash 1017 OpenCode Max is a separate comparison arm from native-harness Max; shared model identity is not shared agent performance. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + DeepSeek V4.1 Flash 1017 · High
DeepSeek Flash 1017 OpenCode High cannot inherit Max's measured quality or resource use. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + MiniMax M3.1 Flash · Max
MiniMax M3.1 Flash OpenCode Max needs disclosed route, exact effort and reproducible outcomes before comparative qualification. Agentic placement additionally requires sustained completion and intervention evidence.
OpenCode + Recursive Flint · rec 4-bit · Default
Recursive Flint rec 4-bit is a precision-specific configuration. Results from a different quantization or provider route cannot establish its coding quality. Agentic placement additionally requires sustained completion and intervention evidence.
Kimi K3.1 Pro + Kimi Code — default effort
Kimi K3.1 Pro Kimi Code default is not the same arm as OpenCode Max or explicit Kimi Code High. Resolve the actual default before comparing. Agentic placement additionally requires sustained completion and intervention evidence.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low
Atomic Agent 0.6.9 Kimi Low needs its own task record; Kimi Code and OpenCode results do not transfer. Agentic placement additionally requires sustained completion and intervention evidence.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max
Atomic Agent 0.6.9 Kimi Max must be compared separately from Low and from other Kimi harnesses. Agentic placement additionally requires sustained completion and intervention evidence.
Codex + GPT-6.1 Terra — provider default
Terra's fresh benchmark page reports explicit effort arms; these cannot rank the registered provider-default Codex configuration. Determine its actual effort and match the usable configuration first. Agentic placement additionally requires sustained completion and intervention evidence.
GitHub Copilot Fusion · Copilot CLI · Standard
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · High
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · Max
The Oct 28 original AA test verified Fusion Max in the Copilot CLI: led terminal tasks, trailed Opus Max overall and on repository questions with a lower estimated cost per task. Organization-specific routing, intervention cost and independent same-configuration work remain open.
Particle-1 · Codex via Vercel AI Gateway · provider default
Gateway access is verified. No matched repository comparison establishes whether token savings survive retries or preserve correctness.
Codex + GPT-6.1 Terra — XHigh
One partially reviewed October 27 BuilderBench run used Terra XHigh in Codex. The comparison used a different harness for Kimi; neither costs nor general performance are settled.
Muse Wonder Full · Muse Code Contributor · High
VulcanBench's Oct 13–17 reconstruction suite tested High in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Extra-high
VulcanBench's Oct 13–17 reconstruction suite tested Extra-high in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Low
VulcanBench's Oct 13–17 reconstruction suite tested Low in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. 18 of 23 tasks fully solved. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4
Resolved 40 of 50 offline SWE-bench Verified instances on one RTX PRO 6000. Same seeded subset has not been run on current podium rivals, so not a head-to-head win.
GPT-6.1 Terra · Codex CLI 0.159.0 · Low
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Medium
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · High
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Max
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
OpenCode + Tencent Hy4 · OpenRouter default
OpenCode can route to production Tencent Hy4. No matched repository run verified at default effort.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Claude Code + Claude Haiku 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Claude Haiku 5.5 MediumUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Sonnet 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol LowUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol MediumUnresolved since Oct 22, 2026 · Priority reassessment
Cursor + Grok 4.7 HighUnresolved since Oct 22, 2026 · Priority reassessment
DeepSeek Harness + V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Antigravity September 2026 · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · HighUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · GitHub Copilot CLI · provider defaultUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot + GPT-6.1 SolUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusionUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion Max · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Mellum2.1 Thinking · vLLM · documented thinking settingsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · High goal modeUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Beam 1.0 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + MiniMax M3.1 Flash · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Mistral Large 4 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Nemotron 3.5 Super VL · HighUnresolved since Oct 22, 2026 · Priority reassessment
Sonnet 5.5 + Replit Agent Power (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Step 5 PreviewUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Recursive Flint · rec 4-bit · DefaultUnresolved since Oct 22, 2026 · Priority reassessment
Kimi K3.1 Pro + Kimi Code — default effortUnresolved since Oct 23, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — LowUnresolved since Oct 25, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — MaxUnresolved since Oct 25, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — provider defaultUnresolved since Oct 26, 2026 · Priority reassessment
Claude Code + Opus 5Unresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-5.6 SolUnresolved since Oct 29, 2026 · Priority reassessment
Cursor + Fable 5.1Unresolved since Oct 29, 2026 · Priority reassessment
Devin + SWE-2Unresolved since Oct 29, 2026 · Priority reassessment
GitHub Copilot + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — XHighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · HighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · MaxUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · StandardUnresolved since Oct 31, 2026
Particle-1 · Codex via Vercel AI Gateway · provider defaultUnresolved since Oct 31, 2026
Muse Wonder Full · Muse Code Contributor · Extra-highUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · HighUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · LowUnresolved since Nov 1, 2026
Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4Unresolved since Nov 2, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · HighUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · LowUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MaxUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MediumUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · XHighUnresolved since Nov 4, 2026
OpenCode + Tencent Hy4 · OpenRouter defaultUnresolved since Nov 4, 2026

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

50 eligible contenders are not covered by the published assessment. Their absence is not a quality verdict.

Review dates and coverage

Assessment published: Sep 26, 2026
Ranking order changed: Sep 26, 2026
Supporting observations: Sep 12, 2026 – Sep 26, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 9 of 59 eligible models

Still needing comparison: Codex + GPT-6 Sol High; Claude Code + Sonnet 5.5 High; Codex + GPT-6.1 Sol High; Codex + GPT-6.1 Sol Medium; GitHub Copilot + GPT-6.1 Sol; GitHub Copilot HydraFusion; GPT-6.1 Sol + Replit Agent Max Mode (recommended effort); Sonnet 5.5 + Replit Agent Power (recommended effort); Step 5 Preview; Claude Code + Claude Haiku 5.5 High; Claude Code + Claude Haiku 5.5 Medium; Codex + GPT-6.1 Sol Low; Cursor + Grok 4.7 High; Mellum2.1 Thinking · vLLM · documented thinking settings; Muse Wonder · Muse Code · provider defaults; Muse Wonder Full · Muse Code · provider defaults; Gemini 4 Argon · Antigravity September 2026 · Medium; Gemini 4 Argon · GitHub Copilot CLI · provider default; Muse Wonder Full · Muse Code · High goal mode; Gemini 4 Argon · Cursor · High; Gemini 4 Argon · Cursor · Medium; GitHub Copilot HydraFusion · October 14; GitHub Copilot HydraFusion Max · October 14; OpenCode + Beam 1.0 · Max; OpenCode + Nemotron 3.5 Super VL · High; OpenCode + Mistral Large 4 · High; DeepSeek Harness + V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · High; OpenCode + MiniMax M3.1 Flash · Max; OpenCode + Recursive Flint · rec 4-bit · Default; Kimi K3.1 Pro + Kimi Code — default effort; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max; Codex + GPT-6.1 Terra — provider default; GitHub Copilot Fusion · Copilot CLI · Standard; GitHub Copilot Fusion · Copilot CLI · High; GitHub Copilot Fusion · Copilot CLI · Max; Particle-1 · Codex via Vercel AI Gateway · provider default; Codex + GPT-6.1 Terra — XHigh; Muse Wonder Full · Muse Code Contributor · High; Muse Wonder Full · Muse Code Contributor · Extra-high; Muse Wonder Full · Muse Code Contributor · Low; Recursive Flint · mini-SWE-agent 2.4.6 · offline NVFP4; GPT-6.1 Terra · Codex CLI 0.159.0 · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Medium; GPT-6.1 Terra · Codex CLI 0.159.0 · High; GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh; GPT-6.1 Terra · Codex CLI 0.159.0 · Max; OpenCode + Tencent Hy4 · OpenRouter default.

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Writing Code

Winner · #1

Claude Code + Opus 5.5 High

Anthropic / Claude Code

Demanding implementation and everyday codebase changes

Model API$4 in $20 out per 1M tokens

* Early evidence — position may changeTested configuration

Why it wins & tradeoffs

An early lead at High effort in a six-project comparison, backed by benchmarks and daily-use reports. Serious failures keep the recommendation provisional.

Price & access: Claude paid plan or API billing; select Opus 5.5 High. Cost per finished task varies by workload.
Watch out for: Keep human review and tests in the loop. Reports include broken workflows; long-running unattended reliability is not established.
Content restrictions: Moderate
Evidence: A six-project test and Artificial Analysis favor Opus over Astra at High effort. Daily-use reports include both useful patches and serious failures; broader replication is missing.
System version: September 2026 Claude Code; exact build undisclosed in external test
Tested access: Claude paid plan or API billing; High effort; default safeguards

Sources and observations (4)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code with Opus 5.5 High leads the six-project aggregate, ahead of Codex Astra High, with lower measured average API cost and task time. Perfect published behavioral subtotal; strong quality grading. Actual evaluation dated September 23. Limited projects and repetitions, automated quality judging, harness build undisclosed; publisher-specific sponsorship unknown. This directly tests High in Claude Code; not a debugging-specialty result.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26 comparison: Opus 5.5 adaptive High with default fallback leads Astra High on Terminal-Bench 4.0 and SciCode, but Astra xhigh remains ahead of Opus High on Terminal-Bench. AA's model evaluation is contextual for Claude Code, not a Claude Code harness score. Intelligence-index cost per task is slightly higher than Astra High despite lower token prices; do not promise universally cheaper completion.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Opus 5.5 High: creator preferred its outputs on all three builds for detail, prompt fidelity and gameplay. Supports creative coding quality; does not establish superiority on repository repair or every coding task. Creator still prefers Astra for everyday refactors/architecture. Scope: generated code/output fidelity across three tasks; auxiliary image tools must not be credited to native model generation.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 Vibe Check: original Claude Code/app tests support daily implementation and concise patches, with some builders switching from Fable/Codex. Contrary evidence includes broken core app flows, uncalled required service and destructive verification behavior. Examples use mixed efforts (including xhigh), so they are contextual for High, not matched High victories. Every received Anthropic pre-launch access; editorial independence asserted by Every, no independent-replication claim.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 23, 2026Provider, pricing & access ↗

Codex + GPT-6 Astra

OpenAI / Codex

Codebase and terminal work with adjustable reasoning effort

* Early evidence — position may changeTested configuration

When to choose Codex + GPT-6 Astra

A close alternative that completed the test's required behaviors but trailed Opus at High effort. Extra-high terminal results are a separate strength.

Model API$10 in $50 out per 1M tokens

Price & access: Codex, ChatGPT Work or API; access and cost depend on plan and effort.
Watch out for: Opus leads the reviewed six-project test at High effort. Astra's stronger terminal results use a different effort setting.
Content restrictions: Low
Evidence: Opus leads the reviewed High-versus-High coding comparisons. Astra leads separate terminal tests at extra-high effort; those settings are not interchangeable.
System version: September 2026 production harness
Tested access: Codex, ChatGPT Work, or API

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot explicitly tests Codex CLI + GPT-6 Astra High and Medium separately. High trails Claude Code + Opus 5.5 High on aggregate quality; both clear the behavior tasks. High is slower and costlier on this suite. High evidence is a configuration-specific observation within the historically broad Astra entry, not evidence about xhigh/max settings.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 26: Astra High trails Opus 5.5 High on terminal and scientific coding, while Astra xhigh retains a terminal advantage over Opus High. AA evaluates a different harness from Codex; settings must stay explicit. Astra retains stronger AutomationBench results. The broad Astra configuration supports choosing xhigh for hard terminal work, not claiming its High setting wins.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Codex + GPT-6 Astra, explicitly High in this comparison: Opus won creator preference on all three creative builds; Astra was faster in reported runs and cheapest on two. Creator still favors Astra for everyday refactors/architecture, a reported workflow preference rather than a measured debugging test. Do not transfer to xhigh or other efforts. Scope: generated code/output fidelity across three tasks; auxiliary image tools must not be credited to native model generation.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review retains Astra for time-bounded work and reports higher raw throughput in one coding experiment, while Opus met more latency budgets. Scope is one experiment with mixed settings, not an Astra High benchmark. September 3 production-rewrite benchmark remains separate and used Astra extra-high.
Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
The software-build task compared the exact Codex + Astra and Claude Code + Fable 5.1 configurations rather than attributing results to naked models.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Claude Code + Fable 5.1

Anthropic / Claude Code

Difficult implementation and careful changes to existing code

* Early evidence — position may changeTested configuration

When to choose Claude Code + Fable 5.1

A strong choice for difficult codebase problems, though newer comparisons favor other models. Tests use different effort settings.

Model API$10 in $50 out per 1M tokens

Price & access: Claude subscription or API billing; heavier usage can be costly.
Watch out for: Newer models lead some implementation tests and may cost less per task.
Content restrictions: Moderate
Evidence: Every still prefers Fable for some hard tasks. Other comparisons use Medium or different effort settings, so they do not establish a uniform defeat.
System version: September 2026 production release
Tested access: Claude Max or API billing

Sources and observations (5)

AI Coding Daily — original coding benchmarkSource record · Observed Sep 26, 2026
September 25 leaderboard snapshot: Claude Code + Fable 5.1 Medium trails Opus 5.5 High and Astra High on the six-project aggregate. The Fable row is Medium: it does not measure Fable High/Max. Retain that scope and use separate evidence for hard-problem frontier capability.
Artificial AnalysisSource record · Observed Sep 26, 2026
AA September 22 original Opus 5.5 analysis: Opus max surpasses Fable 5.1's prior SciCode result and overall index, but this max comparison is contextual and cannot be relabeled Opus High versus Fable High. Fable remains a frontier coding model with premium pricing. Retain the older direct Fable configuration observations alongside this version-specific challenge.
Pat Simmons / AI for Mortals — practical coding comparisonsSource record · Observed Sep 23, 2026
Original comparison VxzdNX6mNSQ and companion AI for Mortals results article; tests September 23, 2026. Same prompts, High effort on all models in Claude Code/Codex, three creative builds (code-only illustration, launch site/film, 3D game). Exact harness builds undisclosed; shared gpt-image-2 where allowed. Single creator's practical comparison; not a BMN reproduction, objective repository benchmark, or separate publisher from the article. Claude Code + Fable 5.1, explicitly High in this comparison: Opus won creator preference across all three creative builds. Fable showed speed but missed some requirements. These are task-level observations, not a universal third-place claim or evidence for other effort settings. Scope: generated code/output fidelity across three tasks; auxiliary image tools must not be credited to native model generation.
Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource record · Observed Sep 22, 2026
Every September 22 Opus 5.5 review: some daily coding moves to Opus for value, but two testers still prefer Fable 5.1 on the hardest coding problems. Ruby comparison used Fable High versus Opus xhigh; do not treat effort as matched. This supports retaining Fable as a demanding-task alternative. Anthropic supplied pre-launch access to Opus; keep that relationship disclosed.
Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
The software-build task compared the exact Codex + Astra and Claude Code + Fable 5.1 configurations rather than attributing results to naked models.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 26, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

What we judge

For turning clear instructions into correct, maintainable code. The focus is the implementation itself.

Functional correctnessFollowing instructionsCode qualityLanguage rangeEfficiency

About this provisional ranking

Based on external tests and hands-on comparisons. Results depend on the model, reasoning effort and coding tools. More testing may change the order; reviews are due every two days.

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (56)

The order is unchanged; some comparisons are still missing.

Codex CLI 0.159.0 plus Terra Max solved 23/23 VulcanBench reconstruction tasks versus 16/23 at XHigh. Small Python repair success at all tested Low–XHigh settings is useful effort guidance, not new generation superiority. Preserve BuilderBench's distinct harness and original observations.

Next comparison: Seek independent same-harness comparisons of new implementation, tests and multi-file revision, with runtime and cost.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Current High-effort implementation remains competitive, but the older Astra agentic/refactoring rationale mixes effort settings and dates. Keep it as a historical recommendation, not proof that every Codex Astra setting defeats newer contenders.
Claude Code + Fable 5.1
Keep Fable 5.1's historical careful-codebase rationale with its original dates. The new effort table uses Medium, so it cannot settle High/Max performance or refresh the old general configuration.
Claude Code + Opus 5
Opus 5 remains the historical baseline; this pass adds no current exact-version advantage over Opus 5.5 or Sol 6.1. Retention is not a new win.
Cursor + Fable 5.1
Cursor + Fable 5.1 needs a Cursor-specific result; Claude Code findings cannot establish this configuration's quality or agent reliability.
Codex + GPT-5.6 Sol
GPT-5.6 Sol is an older Codex baseline. New Sol 6.1 findings are not transferable evidence for this version.
Replit Agent + Automatically Routed
Automatic routing offers an integrated workflow, but the routed stack and cost per completed task must be disclosed before model-specific quality attribution.
GitHub Copilot + Automatically Routed
Copilot Auto's convenient workflow does not identify which model completed each task. Require current routing, intervention and task-completion records.
Devin + SWE-2
Devin + SWE-2 has saved mixed repair/terminal results; the older vendor and selected-suite evidence cannot establish a current across-task lead.
Claude Code + Opus 5.5 High
Opus 5.5 High remains a reasonable implementation shortlist choice: recent short-task and practical-build findings support it, while intervention and sustained-task reliability remain limitations.
Codex + GPT-6 Sol High
GPT-6 Sol High is a separate predecessor to Sol 6.1. Its saved evidence is retained without borrowing successor results.
Claude Code + Sonnet 5.5 High
Sonnet 5.5 High is represented in the current short-task comparison. A credible lower-cost candidate, but not a demonstrated all-workload replacement for the frontier choices.
Codex + GPT-6.1 Sol High
Sol 6.1 High is a credible quality/value shortlist alternative. Current short-task evidence is close to the leaders, with narrow reconstruction support; exact independent replication and broad reliability remain open.
Codex + GPT-6.1 Sol Medium
Sol 6.1 Medium is a sensible speed/value setting to consider for bounded changes. Its weaker-effort result is its own tradeoff, not proof of High-level quality.
GitHub Copilot + GPT-6.1 Sol
Copilot's Sol 6.1 default setting needs Copilot-specific observation. Codex effort results cannot establish its behavior.
GitHub Copilot HydraFusion
September HydraFusion cannot inherit results for the October adaptive or Max releases; record the routing and build used by any comparison.
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)
Replit Max Mode changes harness and effort selection. Codex Sol High findings do not establish this usable configuration's quality or cost.
Sonnet 5.5 + Replit Agent Power (recommended effort)
Replit Power with Sonnet requires its own completion and recovery evidence; Claude Code Sonnet High is a different configuration.
Step 5 Preview
Step 5 Preview's coding eligibility does not establish a reproducible agent/harness configuration or comparative repository performance.
Claude Code + Claude Haiku 5.5 High
Haiku 5.5 High is a separate effort configuration requiring task-specific correctness and recovery evidence; speed or vendor availability alone is insufficient.
Claude Code + Claude Haiku 5.5 Medium
Haiku 5.5 Medium needs its own observed tradeoff against High; neither effort inherits the other's result.
Codex + GPT-6.1 Sol Low
Sol 6.1 Low is worth considering for quick, bounded changes. Initial effort findings suggest weaker quality than High; require review and tests rather than treating lower cost as equal capability.
Cursor + Grok 4.7 High
Cursor + Grok 4.7 High is a usable challenger. Available narrow benchmark signals require artifact, funding and workload checks before a broad editorial order.
Mellum2.1 Thinking · vLLM · documented thinking settings
Mellum2.1 Thinking in vLLM needs actual repository tasks with the documented thinking parameters, not model-level or completion-only evidence.
Muse Wonder · Muse Code · provider defaults
Muse Wonder defaults are distinct from Muse Wonder Full High goal mode. The latter's practical failures are not evidence that this exact configuration fails.
Muse Wonder Full · Muse Code · provider defaults
Muse Wonder Full defaults cannot inherit the High goal-mode result. Default execution, quota behavior and task recovery still require direct observation.
Gemini 4 Argon · Antigravity September 2026 · Medium
Argon in Antigravity Medium needs harness-matched repository outcomes; Cursor, Copilot and bare-model results are not interchangeable.
Gemini 4 Argon · GitHub Copilot CLI · provider default
Argon in Copilot's provider default needs recorded default/routing behavior and actual task outcomes.
Muse Wonder Full · Muse Code · High goal mode
Pat tested High goal mode with Contributor off; VulcanBench High used Contributor on in 1.4.9. Preserve Pat’s interrupted quota and estimated cost. These are not a replication.
Gemini 4 Argon · Cursor · High
Argon Cursor High requires its own exact-effort outcomes and cost/intervention record.
Gemini 4 Argon · Cursor · Medium
Argon Cursor Medium must be assessed independently from High; no effort uplift is inferred.
GitHub Copilot HydraFusion · October 14
October 14 HydraFusion adaptive routing is a materially different stack from September. Require disclosed routed models and repeatable task records.
GitHub Copilot HydraFusion Max · October 14
HydraFusion Max needs evidence for its own maximum-quality routing and usage costs, not automatic inheritance from adaptive mode.
OpenCode + Beam 1.0 · Max
Beam 1.0 OpenCode Max is in the usable field. Current short-task signals do not resolve independent exact-configuration reliability and completion quality.
OpenCode + Nemotron 3.5 Super VL · High
Nemotron 3.5 Super VL OpenCode High needs corroborated repository outcomes with exact routing/build details.
OpenCode + Mistral Large 4 · High
Mistral Large 4 OpenCode High needs task-matched independent replication before a broad placement.
DeepSeek Harness + V4.1 Flash 1017 · Max
DeepSeek V4.1 Flash 1017 in the native harness is distinct from OpenCode. Assess each one's completion, failures and costs separately.
OpenCode + DeepSeek V4.1 Flash 1017 · Max
DeepSeek Flash 1017 OpenCode Max is a separate comparison arm from native-harness Max; shared model identity is not shared agent performance.
OpenCode + DeepSeek V4.1 Flash 1017 · High
DeepSeek Flash 1017 OpenCode High cannot inherit Max's measured quality or resource use.
Claude Code + MiniMax M3.1 Flash (Max, native MiniMax)
Native MiniMax M3.1 Flash Max through Claude Code needs route-specific evidence; third-party or OpenCode runs may behave differently.
OpenCode + MiniMax M3.1 Flash · Max
MiniMax M3.1 Flash OpenCode Max needs disclosed route, exact effort and reproducible outcomes before comparative qualification.
OpenCode + Recursive Flint · rec 4-bit · Default
Recursive Flint rec 4-bit is a precision-specific configuration. Results from a different quantization or provider route cannot establish its coding quality.
Kimi K3.1 Pro + Kimi Code — default effort
Kimi K3.1 Pro Kimi Code default is not the same arm as OpenCode Max or explicit Kimi Code High. Resolve the actual default before comparing.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low
Atomic Agent 0.6.9 Kimi Low needs its own task record; Kimi Code and OpenCode results do not transfer.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max
Atomic Agent 0.6.9 Kimi Max must be compared separately from Low and from other Kimi harnesses.
Codex + GPT-6.1 Terra — provider default
Terra's fresh benchmark page reports explicit effort arms; these cannot rank the registered provider-default Codex configuration. Determine its actual effort and match the usable configuration first.
GitHub Copilot Fusion · Copilot CLI · Standard
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · High
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · Max
The Oct 28 original AA test verified Fusion Max in the Copilot CLI: led terminal tasks, trailed Opus Max overall and on repository questions with a lower estimated cost per task. Organization-specific routing, intervention cost and independent same-configuration work remain open.
Particle-1 · Codex via Vercel AI Gateway · provider default
Gateway access is verified. No matched repository comparison establishes whether token savings survive retries or preserve correctness.
Codex + GPT-6.1 Terra — XHigh
One partially reviewed October 27 BuilderBench run used Terra XHigh in Codex. The comparison used a different harness for Kimi; neither costs nor general performance are settled.
Muse Wonder Full · Muse Code Contributor · High
VulcanBench's Oct 13–17 reconstruction suite tested High in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Extra-high
VulcanBench's Oct 13–17 reconstruction suite tested Extra-high in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Low
VulcanBench's Oct 13–17 reconstruction suite tested Low in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. 18 of 23 tasks fully solved. No equivalent evidence for Contributor-off goal mode or broad independent replication.
GPT-6.1 Terra · Codex CLI 0.159.0 · Low
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Medium
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · High
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Max
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
OpenCode + Tencent Hy4 · OpenRouter default
OpenCode can route to production Tencent Hy4. No matched repository run verified at default effort.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Claude Code + Claude Haiku 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Claude Haiku 5.5 MediumUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + MiniMax M3.1 Flash (Max, native MiniMax)Unresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Sonnet 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol LowUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol MediumUnresolved since Oct 22, 2026 · Priority reassessment
Cursor + Grok 4.7 HighUnresolved since Oct 22, 2026 · Priority reassessment
DeepSeek Harness + V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Antigravity September 2026 · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · HighUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · GitHub Copilot CLI · provider defaultUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot + GPT-6.1 SolUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusionUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion Max · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Mellum2.1 Thinking · vLLM · documented thinking settingsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · High goal modeUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Beam 1.0 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + MiniMax M3.1 Flash · MaxUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Mistral Large 4 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Nemotron 3.5 Super VL · HighUnresolved since Oct 22, 2026 · Priority reassessment
Sonnet 5.5 + Replit Agent Power (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Step 5 PreviewUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Recursive Flint · rec 4-bit · DefaultUnresolved since Oct 22, 2026 · Priority reassessment
Kimi K3.1 Pro + Kimi Code — default effortUnresolved since Oct 23, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — LowUnresolved since Oct 25, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — MaxUnresolved since Oct 25, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — provider defaultUnresolved since Oct 26, 2026 · Priority reassessment
Claude Code + Opus 5Unresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-5.6 SolUnresolved since Oct 29, 2026 · Priority reassessment
Cursor + Fable 5.1Unresolved since Oct 29, 2026 · Priority reassessment
Devin + SWE-2Unresolved since Oct 29, 2026 · Priority reassessment
GitHub Copilot + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — XHighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · HighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · MaxUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · StandardUnresolved since Oct 31, 2026
Particle-1 · Codex via Vercel AI Gateway · provider defaultUnresolved since Oct 31, 2026
Muse Wonder Full · Muse Code Contributor · Extra-highUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · HighUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · LowUnresolved since Nov 1, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · HighUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · LowUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MaxUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MediumUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · XHighUnresolved since Nov 4, 2026
OpenCode + Tencent Hy4 · OpenRouter defaultUnresolved since Nov 4, 2026

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

50 eligible contenders are not covered by the published assessment. Their absence is not a quality verdict.

Review dates and coverage

Assessment published: Sep 26, 2026
Ranking order changed: Sep 26, 2026
Supporting observations: Sep 12, 2026 – Sep 26, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 9 of 59 eligible models

Still needing comparison: Codex + GPT-6 Sol High; Claude Code + Sonnet 5.5 High; Codex + GPT-6.1 Sol High; Codex + GPT-6.1 Sol Medium; GitHub Copilot + GPT-6.1 Sol; GitHub Copilot HydraFusion; GPT-6.1 Sol + Replit Agent Max Mode (recommended effort); Sonnet 5.5 + Replit Agent Power (recommended effort); Step 5 Preview; Claude Code + Claude Haiku 5.5 High; Claude Code + Claude Haiku 5.5 Medium; Codex + GPT-6.1 Sol Low; Cursor + Grok 4.7 High; Mellum2.1 Thinking · vLLM · documented thinking settings; Muse Wonder · Muse Code · provider defaults; Muse Wonder Full · Muse Code · provider defaults; Gemini 4 Argon · Antigravity September 2026 · Medium; Gemini 4 Argon · GitHub Copilot CLI · provider default; Muse Wonder Full · Muse Code · High goal mode; Gemini 4 Argon · Cursor · High; Gemini 4 Argon · Cursor · Medium; GitHub Copilot HydraFusion · October 14; GitHub Copilot HydraFusion Max · October 14; OpenCode + Beam 1.0 · Max; OpenCode + Nemotron 3.5 Super VL · High; OpenCode + Mistral Large 4 · High; DeepSeek Harness + V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · High; Claude Code + MiniMax M3.1 Flash (Max, native MiniMax); OpenCode + MiniMax M3.1 Flash · Max; OpenCode + Recursive Flint · rec 4-bit · Default; Kimi K3.1 Pro + Kimi Code — default effort; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max; Codex + GPT-6.1 Terra — provider default; GitHub Copilot Fusion · Copilot CLI · Standard; GitHub Copilot Fusion · Copilot CLI · High; GitHub Copilot Fusion · Copilot CLI · Max; Particle-1 · Codex via Vercel AI Gateway · provider default; Codex + GPT-6.1 Terra — XHigh; Muse Wonder Full · Muse Code Contributor · High; Muse Wonder Full · Muse Code Contributor · Extra-high; Muse Wonder Full · Muse Code Contributor · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Medium; GPT-6.1 Terra · Codex CLI 0.159.0 · High; GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh; GPT-6.1 Terra · Codex CLI 0.159.0 · Max; OpenCode + Tencent Hy4 · OpenRouter default.

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Fixing & Improving Code

Winner · #1

Claude Code + Fable 5.1

Anthropic / Claude Code

Careful fixes and improvements to large codebases

Model API$10 in $50 out per 1M tokens

Weighted score: 92.2Tested configuration

Why it wins & tradeoffs

Leads for careful codebase judgment and disciplined changes.

Price & access: Claude Max or API billing; heavy usage can become expensive.
Watch out for: Can be slower and more expensive per completed task.
Content restrictions: Moderate
Evidence: Current independent benchmarks and practitioner testing support frontier quality and strong product judgment, with higher cost and slower task completion.
System version: September 2026 production release
Tested access: Claude Max or API billing

Sources and observations (1)

Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource page · Observed Sep 3, 2026
Every's current hands-on tests favored Fable 5.1's product judgment, while documenting specification-overrun risk.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Codex + GPT-6 Astra

OpenAI / Codex

Fast codebase fixes, rewrites and terminal work

Weighted score: 90.6Tested configuration

When to choose Codex + GPT-6 Astra

Strong results on production rewrites and efficiency, but can make broader changes than wanted.

Model API$10 in $50 out per 1M tokens

Price & access: Codex, ChatGPT Work, or API; frontier access remains plan-dependent.
Watch out for: Review the scope of its changes; access also depends on your plan.
Content restrictions: Low
Evidence: Artificial Analysis, terminal tests, Every and practitioner comparisons support strong performance, speed and cost efficiency.
System version: September 2026 production harness
Tested access: Codex, ChatGPT Work, or API

Sources and observations (2)

Every — current coding-agent Vibe Checks and Senior Engineer BenchmarkSource page · Observed Sep 3, 2026
Astra's 71/100 production-code rewrite materially exceeded both Fable 5.1 runs, but Every documented overbuilding and stopping-judgment weaknesses.
Artificial AnalysisReferenced source page · Observed Sep 12, 2026
Referenced in the published evaluation or explanation. A specific source observation is not attached here; this link is context, not a substitute for a documented comparison.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Claude Code + Opus 5

Anthropic / Claude Code

Root-cause diagnosis and careful code changes

Weighted score: 88.8Tested configuration

Supporting source links for this published explanation are still being reconciled.

When to choose Claude Code + Opus 5

A dependable third choice for diagnosing root causes and improving existing code.

Model API$5 in $25 out per 1M tokens

Price & access: Claude subscription or API billing.
Watch out for: Newer alternatives may perform better or cost less.
Content restrictions: Moderate
Evidence: Independent repository and coding-agent evidence supports mature, dependable engineering performance.
System version: September 2026 production release
Tested access: Claude Max or API billing

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 7, 2026Provider, pricing & access ↗

Also excellent

Cursor + Fable 5.1

Cursor / Anthropic

A close alternative to third place (0.2 weighted points apart). This gap does not establish a meaningful quality difference.

Model API$10 in $50 out per 1M tokens

Weighted score: 88.6

What we judge

For finding the cause of a bug or improving an existing project without breaking what already works.

DiagnosisFix correctnessRegression avoidanceRefactoring qualityExplanation

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (48)

The order is unchanged; some comparisons are still missing.

Terra now has direct repair evidence, not just repository reconstruction: Max found 40.7/105 planted bugs on average versus 11.3 at XHigh in Bug Hunt Bench's three-repeat runs. All Low–XHigh efforts passed NerdsChalk's small repair, where Max was untested. Use lower effort for bounded fixes; consider Max for difficult searches. Whole-field superiority remains unestablished.

Next comparison: Replicate larger multi-file fixes against incumbents with hidden regressions, exact CLI/effort, runtime and measured cost.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Current High-effort implementation remains competitive, but the older Astra agentic/refactoring rationale mixes effort settings and dates. Keep it as a historical recommendation, not proof that every Codex Astra setting defeats newer contenders. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Claude Code + Fable 5.1
Keep Fable 5.1's historical careful-codebase rationale with its original dates. The new effort table uses Medium, so it cannot settle High/Max performance or refresh the old general configuration. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Claude Code + Opus 5
Opus 5 remains the historical baseline; this pass adds no current exact-version advantage over Opus 5.5 or Sol 6.1. Retention is not a new win. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Cursor + Fable 5.1
Cursor + Fable 5.1 needs a Cursor-specific result; Claude Code findings cannot establish this configuration's quality or agent reliability. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Codex + GPT-5.6 Sol
GPT-5.6 Sol is an older Codex baseline. New Sol 6.1 findings are not transferable evidence for this version. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Replit Agent + Automatically Routed
Automatic routing offers an integrated workflow, but the routed stack and cost per completed task must be disclosed before model-specific quality attribution. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot + Automatically Routed
Copilot Auto's convenient workflow does not identify which model completed each task. Require current routing, intervention and task-completion records. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Devin + SWE-2
Devin + SWE-2 has saved mixed repair/terminal results; the older vendor and selected-suite evidence cannot establish a current across-task lead. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Claude Code + Sonnet 5.5 High
Sonnet 5.5 High is represented in the current short-task comparison. A credible lower-cost candidate, but not a demonstrated all-workload replacement for the frontier choices. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Codex + GPT-6.1 Sol High
Sol 6.1 High is a credible quality/value shortlist alternative. Current short-task evidence is close to the leaders, with narrow reconstruction support; exact independent replication and broad reliability remain open. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Codex + GPT-6.1 Sol Medium
Sol 6.1 Medium is a sensible speed/value setting to consider for bounded changes. Its weaker-effort result is its own tradeoff, not proof of High-level quality. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot + GPT-6.1 Sol
Copilot's Sol 6.1 default setting needs Copilot-specific observation. Codex effort results cannot establish its behavior. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot HydraFusion
September HydraFusion cannot inherit results for the October adaptive or Max releases; record the routing and build used by any comparison. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)
Replit Max Mode changes harness and effort selection. Codex Sol High findings do not establish this usable configuration's quality or cost. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Sonnet 5.5 + Replit Agent Power (recommended effort)
Replit Power with Sonnet requires its own completion and recovery evidence; Claude Code Sonnet High is a different configuration. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Step 5 Preview
Step 5 Preview's coding eligibility does not establish a reproducible agent/harness configuration or comparative repository performance. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Claude Code + Claude Haiku 5.5 High
Haiku 5.5 High is a separate effort configuration requiring task-specific correctness and recovery evidence; speed or vendor availability alone is insufficient. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Claude Code + Claude Haiku 5.5 Medium
Haiku 5.5 Medium needs its own observed tradeoff against High; neither effort inherits the other's result. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Codex + GPT-6.1 Sol Low
Sol 6.1 Low is worth considering for quick, bounded changes. Initial effort findings suggest weaker quality than High; require review and tests rather than treating lower cost as equal capability. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Mellum2.1 Thinking · vLLM · documented thinking settings
Mellum2.1 Thinking in vLLM needs actual repository tasks with the documented thinking parameters, not model-level or completion-only evidence. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Muse Wonder · Muse Code · provider defaults
Muse Wonder defaults are distinct from Muse Wonder Full High goal mode. The latter's practical failures are not evidence that this exact configuration fails. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Muse Wonder Full · Muse Code · provider defaults
Muse Wonder Full defaults cannot inherit the High goal-mode result. Default execution, quota behavior and task recovery still require direct observation. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Gemini 4 Argon · Antigravity September 2026 · Medium
Argon in Antigravity Medium needs harness-matched repository outcomes; Cursor, Copilot and bare-model results are not interchangeable. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Gemini 4 Argon · GitHub Copilot CLI · provider default
Argon in Copilot's provider default needs recorded default/routing behavior and actual task outcomes. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Muse Wonder Full · Muse Code · High goal mode
Pat tested High goal mode with Contributor off; VulcanBench High used Contributor on in 1.4.9. Preserve Pat’s interrupted quota and estimated cost. These are not a replication.
Gemini 4 Argon · Cursor · High
Argon Cursor High requires its own exact-effort outcomes and cost/intervention record. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Gemini 4 Argon · Cursor · Medium
Argon Cursor Medium must be assessed independently from High; no effort uplift is inferred. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot HydraFusion · October 14
October 14 HydraFusion adaptive routing is a materially different stack from September. Require disclosed routed models and repeatable task records. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot HydraFusion Max · October 14
HydraFusion Max needs evidence for its own maximum-quality routing and usage costs, not automatic inheritance from adaptive mode. Repair/refactoring placement additionally requires regression-conscious patch evidence.
DeepSeek Harness + V4.1 Flash 1017 · Max
DeepSeek V4.1 Flash 1017 in the native harness is distinct from OpenCode. Assess each one's completion, failures and costs separately. Repair/refactoring placement additionally requires regression-conscious patch evidence.
OpenCode + DeepSeek V4.1 Flash 1017 · Max
DeepSeek Flash 1017 OpenCode Max is a separate comparison arm from native-harness Max; shared model identity is not shared agent performance. Repair/refactoring placement additionally requires regression-conscious patch evidence.
OpenCode + DeepSeek V4.1 Flash 1017 · High
DeepSeek Flash 1017 OpenCode High cannot inherit Max's measured quality or resource use. Repair/refactoring placement additionally requires regression-conscious patch evidence.
OpenCode + Recursive Flint · rec 4-bit · Default
Recursive Flint rec 4-bit is a precision-specific configuration. Results from a different quantization or provider route cannot establish its coding quality. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Kimi K3.1 Pro + Kimi Code — default effort
Kimi K3.1 Pro Kimi Code default is not the same arm as OpenCode Max or explicit Kimi Code High. Resolve the actual default before comparing. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low
Atomic Agent 0.6.9 Kimi Low needs its own task record; Kimi Code and OpenCode results do not transfer. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max
Atomic Agent 0.6.9 Kimi Max must be compared separately from Low and from other Kimi harnesses. Repair/refactoring placement additionally requires regression-conscious patch evidence.
Codex + GPT-6.1 Terra — provider default
Terra's fresh benchmark page reports explicit effort arms; these cannot rank the registered provider-default Codex configuration. Determine its actual effort and match the usable configuration first. Repair/refactoring placement additionally requires regression-conscious patch evidence.
GitHub Copilot Fusion · Copilot CLI · Standard
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · High
This GA CLI effort is available, but the selected models depend on organization policy. No matched independent trial with that pool establishes quality.
GitHub Copilot Fusion · Copilot CLI · Max
The Oct 28 original AA test verified Fusion Max in the Copilot CLI: led terminal tasks, trailed Opus Max overall and on repository questions with a lower estimated cost per task. Organization-specific routing, intervention cost and independent same-configuration work remain open.
Particle-1 · Codex via Vercel AI Gateway · provider default
Gateway access is verified. No matched repository comparison establishes whether token savings survive retries or preserve correctness.
Codex + GPT-6.1 Terra — XHigh
One partially reviewed October 27 BuilderBench run used Terra XHigh in Codex. The comparison used a different harness for Kimi; neither costs nor general performance are settled.
Muse Wonder Full · Muse Code Contributor · High
VulcanBench's Oct 13–17 reconstruction suite tested High in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Extra-high
VulcanBench's Oct 13–17 reconstruction suite tested Extra-high in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. One three-hour timeout; judging only finished work hides the failure. No equivalent evidence for Contributor-off goal mode or broad independent replication.
Muse Wonder Full · Muse Code Contributor · Low
VulcanBench's Oct 13–17 reconstruction suite tested Low in Muse Code Contributor mode, a subscription with quota interruptions and estimated cost. 18 of 23 tasks fully solved. No equivalent evidence for Contributor-off goal mode or broad independent replication.
GPT-6.1 Terra · Codex CLI 0.159.0 · Low
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Medium
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · High
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
GPT-6.1 Terra · Codex CLI 0.159.0 · Max
Exact Codex CLI0.159.0 Terra effort retained. October26–27 Bug Hunt three-repeat means: Low6, Medium7.3, High9.7, XHigh11.3, Max40.7 of105 bugs. October28 NerdsChalk small Python repair saturated Low–XHigh; Max untested there. Different tasks and original dates; no broad incumbent superiority inferred.
OpenCode + Tencent Hy4 · OpenRouter default
OpenCode can use the production Hy4 route. No matched repository trial verified for this agent and default effort.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Claude Code + Claude Haiku 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Claude Haiku 5.5 MediumUnresolved since Oct 22, 2026 · Priority reassessment
Claude Code + Sonnet 5.5 HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol HighUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol LowUnresolved since Oct 22, 2026 · Priority reassessment
Codex + GPT-6.1 Sol MediumUnresolved since Oct 22, 2026 · Priority reassessment
DeepSeek Harness + V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Antigravity September 2026 · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · HighUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · Cursor · MediumUnresolved since Oct 22, 2026 · Priority reassessment
Gemini 4 Argon · GitHub Copilot CLI · provider defaultUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot + GPT-6.1 SolUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusionUnresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GitHub Copilot HydraFusion Max · October 14Unresolved since Oct 22, 2026 · Priority reassessment
GPT-6.1 Sol + Replit Agent Max Mode (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Mellum2.1 Thinking · vLLM · documented thinking settingsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · High goal modeUnresolved since Oct 22, 2026 · Priority reassessment
Muse Wonder Full · Muse Code · provider defaultsUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · HighUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + DeepSeek V4.1 Flash 1017 · MaxUnresolved since Oct 22, 2026 · Priority reassessment
Sonnet 5.5 + Replit Agent Power (recommended effort)Unresolved since Oct 22, 2026 · Priority reassessment
Step 5 PreviewUnresolved since Oct 22, 2026 · Priority reassessment
OpenCode + Recursive Flint · rec 4-bit · DefaultUnresolved since Oct 22, 2026 · Priority reassessment
Kimi K3.1 Pro + Kimi Code — default effortUnresolved since Oct 23, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — LowUnresolved since Oct 25, 2026 · Priority reassessment
Kimi K3.1 Pro + Atomic Agent 0.6.9 — MaxUnresolved since Oct 25, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — provider defaultUnresolved since Oct 26, 2026 · Priority reassessment
Codex + GPT-5.6 SolUnresolved since Oct 29, 2026 · Priority reassessment
Cursor + Fable 5.1Unresolved since Oct 29, 2026 · Priority reassessment
Devin + SWE-2Unresolved since Oct 29, 2026 · Priority reassessment
GitHub Copilot + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 29, 2026 · Priority reassessment
Codex + GPT-6.1 Terra — XHighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · HighUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · MaxUnresolved since Oct 31, 2026
GitHub Copilot Fusion · Copilot CLI · StandardUnresolved since Oct 31, 2026
Particle-1 · Codex via Vercel AI Gateway · provider defaultUnresolved since Oct 31, 2026
Muse Wonder Full · Muse Code Contributor · Extra-highUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · HighUnresolved since Nov 1, 2026
Muse Wonder Full · Muse Code Contributor · LowUnresolved since Nov 1, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · HighUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · LowUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MaxUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · MediumUnresolved since Nov 4, 2026
GPT-6.1 Terra · Codex CLI 0.159.0 · XHighUnresolved since Nov 4, 2026
OpenCode + Tencent Hy4 · OpenRouter defaultUnresolved since Nov 4, 2026

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

43 eligible contenders are not covered by the published assessment. Their absence is not a quality verdict.

Review dates and coverage

Assessment published: Sep 12, 2026
Ranking order changed: Not recorded
Supporting observations: Sep 12, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 8 of 51 eligible models

Still needing comparison: Claude Code + Sonnet 5.5 High; Codex + GPT-6.1 Sol High; Codex + GPT-6.1 Sol Medium; GitHub Copilot + GPT-6.1 Sol; GitHub Copilot HydraFusion; GPT-6.1 Sol + Replit Agent Max Mode (recommended effort); Sonnet 5.5 + Replit Agent Power (recommended effort); Step 5 Preview; Claude Code + Claude Haiku 5.5 High; Claude Code + Claude Haiku 5.5 Medium; Codex + GPT-6.1 Sol Low; Mellum2.1 Thinking · vLLM · documented thinking settings; Muse Wonder · Muse Code · provider defaults; Muse Wonder Full · Muse Code · provider defaults; Gemini 4 Argon · Antigravity September 2026 · Medium; Gemini 4 Argon · GitHub Copilot CLI · provider default; Muse Wonder Full · Muse Code · High goal mode; Gemini 4 Argon · Cursor · High; Gemini 4 Argon · Cursor · Medium; GitHub Copilot HydraFusion · October 14; GitHub Copilot HydraFusion Max · October 14; DeepSeek Harness + V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · Max; OpenCode + DeepSeek V4.1 Flash 1017 · High; OpenCode + Recursive Flint · rec 4-bit · Default; Kimi K3.1 Pro + Kimi Code — default effort; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Low; Kimi K3.1 Pro + Atomic Agent 0.6.9 — Max; Codex + GPT-6.1 Terra — provider default; GitHub Copilot Fusion · Copilot CLI · Standard; GitHub Copilot Fusion · Copilot CLI · High; GitHub Copilot Fusion · Copilot CLI · Max; Particle-1 · Codex via Vercel AI Gateway · provider default; Codex + GPT-6.1 Terra — XHigh; Muse Wonder Full · Muse Code Contributor · High; Muse Wonder Full · Muse Code Contributor · Extra-high; Muse Wonder Full · Muse Code Contributor · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Low; GPT-6.1 Terra · Codex CLI 0.159.0 · Medium; GPT-6.1 Terra · Codex CLI 0.159.0 · High; GPT-6.1 Terra · Codex CLI 0.159.0 · XHigh; GPT-6.1 Terra · Codex CLI 0.159.0 · Max; OpenCode + Tencent Hy4 · OpenRouter default.

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Build something

Website & app builders

What makes a good builder? →

Practitioner reviews

Best Website Builder

Winner · #1

Codex + GPT-6 Astra

OpenAI / Codex

Distinctive marketing sites with code ownership

Model API$10 in $50 out per 1M tokens

Weighted score: 92.3Provisional: evidence still developingTested configuration

Why it wins & tradeoffs

Our provisional design leader for polished custom sites, with strong layouts, motion and brand fit.

Price & access: Codex, ChatGPT Work, and API; API starts at $10/M input and $50/M output
Watch out for: Maintaining the site may require coding help; this is not a visual editor.
Content restrictions: Low
Evidence: Nate Herk's reference-site test and independent design tests support the lead. More practitioner comparisons are needed.
System version: September 2026 production harness
Tested access: Codex, ChatGPT Work, or API

Sources and observations (1)

Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
The reference-website test qualifies for Website Builder because both configurations attempted the same site-oriented task; neither result is treated as Mobile App evidence.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

Framer

Marketing sites and portfolios

Weighted score: 91.0Provisional: evidence still developingTested configurationPlatform-selected AI

Supporting source links for this published explanation are still being reconciled.

When to choose Framer

A provisional choice for polished sites with a visual editor and responsive design controls.

Price & access: Free entry tier with paid hosting and collaboration plans
Watch out for: Limited fit for complex app backends or unrestricted code ownership.
Content restrictions: Low
Evidence: Specialist reviews and verified features support a top-three place. More practitioner comparisons are needed.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by Framer. The exact model—or combination of models—may change as the service evolves.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

Claude Code + Fable 5.1

Anthropic / Claude Code

Custom sites built through several rounds of feedback

Weighted score: 90.5Provisional: evidence still developingTested configuration

When to choose Claude Code + Fable 5.1

Produces polished custom sites after revision; needed more rounds than Astra in the reviewed comparison.

Model API$10 in $50 out per 1M tokens

Price & access: Claude Max or API billing
Watch out for: Expect revisions rather than a polished first attempt. There is no visual no-code editor.
Content restrictions: Moderate
Evidence: Nate Herk tested Claude Code with Fable 5.1 on the same reference-site task. More independent comparisons are needed.
System version: September 2026 production release
Tested access: Claude Max or API billing

Sources and observations (1)

Nate Herk — GPT-6 Astra vs Fable 5.1 practical comparisonSource page · Observed Sep 6, 2026
The reference-website test qualifies for Website Builder because both configurations attempted the same site-oriented task; neither result is treated as Mobile App evidence.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 6, 2026Provider, pricing & access ↗

What we judge

Create a portfolio, business homepage, or publication. We look for polished design, easy editing, and a reliable way to publish.

Practitioner reviewsFinished websitesOwnership & publishingEditing & recoveryEase & running costs

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (7)

The order is unchanged; some comparisons are still missing.

Pat Simmons's September23 rebuild favored Fable5.1's readability while missing a 3D effect; the models chose different sites. Framer/Webflow updates document current features, not a matched client handoff. This remains evidence for initial output tradeoffs rather than a new overall winner.

Next comparison: Compare a common responsive-site brief, publishing, a CMS change and nontechnical client handoff.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Codex + GPT-6 Astra: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Claude Code + Fable 5.1
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Claude Code + Fable 5.1: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Replit Agent + Automatically Routed
Replit Agent + Automatically Routed: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Lovable + Current Model Stack
Lovable + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Bolt + Current Model Stack
Bolt + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Base44 + Current Model Stack
Base44 + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Framer + Current Model Stack
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Framer + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Webflow + Current Model Stack
Webflow + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Wix Studio + Current Model Stack
Wix Studio + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Squarespace + Current Model Stack
Squarespace + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Base44 + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Bolt + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Lovable + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 21, 2026 · Priority reassessment
Squarespace + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Webflow + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Wix Studio + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

Review dates and coverage

Assessment published: Sep 7, 2026
Ranking order changed: Not recorded
Supporting observations: Sep 7, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 10 of 10 eligible models

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Practitioner reviews

Best Web App Builder

Winner · #1

Lovable

Complete web-app prototypes with sign-in, data and GitHub export

Weighted score: 97.0Tested configurationPlatform-selected AI

Why it wins & tradeoffs

Three practitioner reviews put Lovable first, with evidence of working apps and code ownership.

Watch out for: Limited framework choice and fine-grained manual control.
Content restrictions: Low
Evidence: Three reviews rank it first; two are independent, and one shows a working, deployed project tracker.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by Lovable. The exact model—or combination of models—may change as the service evolves.

Sources and observations (3)

LOW/CODE Agency — 2026 no-code builder comparisonSource page · Observed Sep 3, 2026
LOW/CODE's explicitly numbered vibe-coding section places Lovable first and calls it the best AI app builder for nontechnical founders, based on recent internal testing across real workflows.
Vibe Coding Academy — 2026 AI app builder guideSource page · Observed May 1, 2026
Vibe Coding Academy names Lovable the strongest all-around AI web app builder and demonstrates a deployed project tracker with authentication, database-backed projects and tasks, and iterative dashboard changes.
Till Freitag — Base44 vs Lovable agency comparisonSource page · Observed Mar 17, 2026
After using both builders on several agency projects and comparing a CRM dashboard workflow, Till Freitag prefers Lovable for production and client work because of code and backend portability; its Lovable services create a material commercial-interest discount.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

Bolt

StackBlitz

Quick web-app prototypes with code ownership

Weighted score: 49.3Tested configurationPlatform-selected AI

When to choose Bolt

A strong runner-up for fast revisions and exportable code.

Watch out for: Setting up the backend may take more work than with an all-in-one builder.
Content restrictions: Low
Evidence: LOW/CODE ranks Bolt second; provider documentation supports code ownership and publishing.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by Bolt. The exact model—or combination of models—may change as the service evolves.

Sources and observations (1)

LOW/CODE Agency — 2026 no-code builder comparisonSource page · Observed Sep 3, 2026
LOW/CODE places Bolt second, favoring it for developers who want full code control and exportability.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

Base44

Wix / Base44

Internal tools and data-heavy business apps

Weighted score: 37.8Tested configurationPlatform-selected AI

When to choose Base44

An approachable choice for internal tools, with less freedom to move the backend elsewhere.

Watch out for: Moving the backend to another platform may be difficult.
Content restrictions: Moderate
Evidence: LOW/CODE ranks Base44 third. Practitioner reviews praise ease of use but flag portability limits.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by Base44. The exact model—or combination of models—may change as the service evolves.

Sources and observations (1)

LOW/CODE Agency — 2026 no-code builder comparisonSource page · Observed Sep 3, 2026
LOW/CODE places Base44 third, favoring it for backend-heavy internal tools while noting its weaker frontend customization.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 12, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

What we judge

Build something people use in a browser, such as a booking system, CRM, or writing tool—with accounts, saved data, and working features.

Practitioner reviewsWorking productsOwnership & publishingEase & revisionRecovery & running costs

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (8)

The order is unchanged; some comparisons are still missing.

The verified product routes remain usable, but launch features do not establish an easier production handoff. LOW/CODE's concrete Lovable checklist requires deployed auth, two-account isolation and persistent data. No newly inspected matched Replit/Lovable deployment result closes that comparison.

Next comparison: Compare the same deployed app, two-account data isolation, natural-language revision and recovery without expert intervention.

Latest model assessments

These findings add context without changing ranks or scores.

Codex + GPT-6 Astra
Codex + GPT-6 Astra: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Claude Code + Fable 5.1
Claude Code + Fable 5.1: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Cursor + Fable 5.1
Cursor + Fable 5.1: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Replit Agent + Automatically Routed
Replit Agent + Automatically Routed: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Lovable + Current Model Stack
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Lovable + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Bolt + Current Model Stack
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Bolt + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Base44 + Current Model Stack
Existing published placement preserved; this limited review does not claim fresh full comparative confirmation. Base44 + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
FlutterFlow + Current Model Stack
FlutterFlow + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Draftbit + Current Model Stack
Draftbit + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Adalo + Current Model Stack
Adalo + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.
Google AI Studio + Current Model Stack
Google AI Studio + Current Model Stack: Verified for the category's required workflow and comparison scope. Current exact-category comparative coverage remains insufficient; absence is not a loss.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

Adalo + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Claude Code + Fable 5.1Unresolved since Oct 21, 2026 · Priority reassessment
Codex + GPT-6 AstraUnresolved since Oct 21, 2026 · Priority reassessment
Cursor + Fable 5.1Unresolved since Oct 21, 2026 · Priority reassessment
Draftbit + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
FlutterFlow + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Google AI Studio + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Replit Agent + Automatically RoutedUnresolved since Oct 21, 2026 · Priority reassessment

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

Review dates and coverage

Assessment published: Sep 12, 2026
Ranking order changed: Not recorded
Supporting observations: Sep 12, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 11 of 11 eligible models

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Practitioner reviews

Best Mobile App Builder

Winner · #1

FlutterFlow

iOS and Android apps intended for the app stores

Weighted score: 89.7Provisional: evidence still developingTested configurationPlatform-selected AI

Why it wins & tradeoffs

Our provisional first choice for visual app building, with testing, code export and app-store publishing.

Price & access: Paid FlutterFlow plan required for key export and deployment workflows
Watch out for: You will need to learn the visual editor; chat alone is not enough.
Content restrictions: Low
Evidence: Publishing tools are verified. More independent comparisons building the same app are needed.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by FlutterFlow. The exact model—or combination of models—may change as the service evolves.

Sources and observations (3)

LOW/CODE Agency — 2026 no-code builder comparisonSource page · Observed Sep 3, 2026
The agency identifies FlutterFlow as its strongest fit for cross-platform native apps needing source ownership, based on recent hands-on evaluation and production experience. The article does not publish a category-specific ranked top-three ballot.
The App Builder Report — Rork vs FlutterFlowSource page · Observed Aug 14, 2026
The comparison explicitly judges FlutterFlow more capable than Rork for repeat building, custom logic, code ownership, and mature store deployment. FlutterFlow was assessed from documentation and independent coverage rather than the author's direct use; the publisher's Rork affiliate relationship discounts the verdict.
Unite.AI — Superapp hands-on reviewSource page · Observed Jul 31, 2026
In the same hands-on review, FlutterFlow is characterized as the more mature choice for cross-platform control and customization. The comparison is useful but affiliate-discounted and is not a complete mobile top-three ballot.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

Draftbit

Visual mobile development with React Native code export

Weighted score: 85.3Provisional: evidence still developingTested configurationPlatform-selected AI

When to choose Draftbit

A provisional runner-up for visual editing with React Native, Expo and exportable code.

Price & access: Paid plans unlock production collaboration and export workflows
Watch out for: Publishing still takes setup; the no-code community is smaller than some alternatives.
Content restrictions: Low
Evidence: Verified tools and mobile-builder comparisons support inclusion. More comparative testing is needed.
System version: September 2026 production release
Tested access: Current public paid plan
AI model: Selected automatically by Draftbit. The exact model—or combination of models—may change as the service evolves.

Sources and observations (1)

LOW/CODE Agency — 2026 no-code builder comparisonSource page · Observed Sep 3, 2026
The agency identifies Draftbit as a strong React Native option for semi-technical teams prioritizing source ownership. The article does not publish a category-specific ranked top-three ballot.

Review dates may differ from publication dates. Some links lead to a source page rather than an individual test.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 12, 2026Provider, pricing & access ↗

Replit Agent

Replit

One workspace for coding, data and deployment

Weighted score: 85.2Provisional: evidence still developingTested configurationPlatform-selected AI

Supporting source links for this published explanation are still being reconciled.

When to choose Replit Agent

A provisional choice for building web and mobile apps in one workspace.

Price & access: Replit Core or Teams with usage-based agent costs
Watch out for: Agent usage costs can be hard to predict, and initial designs may need polishing.
Content restrictions: Low
Evidence: Verified features and builder comparisons support its range. Independent practitioner agreement is still incomplete.
System version: September 2026 production release
Tested access: Replit Core or Teams
AI model: Selected automatically by Replit Agent. The exact model—or combination of models—may change as the service evolves.

Ranking last verified: Sep 10, 2026 · Configuration tested: Sep 7, 2026Provider, pricing & access ↗

What we judge

Create an installable iPhone or Android app. Device testing, code ownership, and a practical path to app stores matter here.

Practitioner reviewsWorking mobile appsCode ownershipApp-store publishingEase, recovery & value

Category reviewed: Nov 5, 2026Priority comparisons overdue

Review details & missing comparisons (8)

The order is unchanged; some comparisons are still missing.

Rork Lab's release account documents fixing sign-in and a privacy URL before App Store acceptance on an Expo/EAS workflow. Today's Rork and Dreamflow documentation confirms available publishing routes, but the old account does not validate the current native stack. Retain provisional choices pending on-device comparisons.

Next comparison: Compare current named native builder workflows on a real device, including failed-build recovery and store submission.

Latest model assessments

These findings add context without changing ranks or scores.

Replit Agent + Automatically Routed
Replit remains an integrated general-purpose option with a historical mobile placement. This pass does not establish that its current automatically routed stack beats newer mobile specialists on a real phone or store deployment.
FlutterFlow + Current Model Stack
FlutterFlow is still a reasonable option for visual control of cross-platform apps if you accept its learning curve. The September 24 release adds navigation visibility, subscription controls and deployment-log recovery. Check existing projects against GenUI redeployment and legacy Firebase retirement notices; these are operational facts, not a new quality victory.
Draftbit + Current Model Stack
Draftbit remains relevant for visual editing plus source-code work. Important ownership caveat: its current docs say adding visual tooling to imported/connected Expo, Vite or Astro projects blocks outside builds until that setup is removed. Evaluate recovery/export for your project type; this does not mean all Draftbit export is broken.
Adalo + Current Model Stack
Adalo remains in the field, but the current targeted material adds no matched on-device result supporting a new placement. Its omission from today's shortlist is an evidence limitation, not a fresh quality defeat.
Google AI Studio + Current Model Stack
Google AI Studio's historical eligibility is retained. General app-generation capability is not proof of a complete native mobile release workflow; require project-specific device, signing, publishing and maintenance outcomes.
Rork + Current Model Stack
Rork deserves an explicit early shortlist for prompt-driven native app creation. Current docs describe Swift iOS and Kotlin Android, so older Expo-era verdicts are not current quality evidence. Practitioner reports support rapid iteration but also developer intervention for polish and operations. Its ranking remains provisional pending current nonexpert, on-device comparison.
Superapp + Current Model Stack
Superapp is a plausible Apple-focused shortlist choice for Swift apps and assisted submission. The July hands-on habit-tracker review is encouraging prototype evidence, not a fresh matched shipping test; affiliate disclosure and Apple-only mobile scope matter. Current automated marketing and low-rejection claims are vendor claims.
Vibecode + Current Model Stack
Vibecode belongs in the active shortlist for users choosing native iOS or cross-platform Expo workflows. Its own October 10 documentation describes export and iOS submission, with a different Android route. The vendor's comparison against Rork is not an independent ballot or proof of superior quality.
VULK + Current Model Stack
VULK is a relevant Flutter/export contender with backend and release scaffolding. Its launch guide correctly requires real-phone persistence and permissions checks. That is a useful evaluation plan, not evidence that a finished app actually passed it.
a0.dev + Current Model Stack
a0.dev remains a relevant mobile candidate with documented publishing and agent guides. Device execution, recovery and store acceptance still need an original practical comparison before a quality placement.
Dreamflow + Current Model Stack
Dreamflow is a plausible alternative for users wanting AI, visual editing and Flutter code together. Its source-code workflow is a meaningful ownership option, but device/store documentation alone does not establish a comparative shipping win.

Limits of this review

Exact configuration, task coverage and independent corroboration limitations remain; existing ranks are provisional and unchanged.

Models needing more comparison evidence

These models remain under review. Missing evidence does not mean they performed worse.

a0.dev + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Adalo + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Dreamflow + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Google AI Studio + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Rork + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Superapp + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
Vibecode + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment
VULK + Current Model StackUnresolved since Oct 21, 2026 · Priority reassessment

We need relevant comparisons with settings and commercial ties disclosed. Rechecking a source alone does not settle quality.

How we handle missing comparisons →

6 eligible contenders are not covered by the published assessment. Their absence is not a quality verdict.

Review dates and coverage

Assessment published: Sep 7, 2026
Ranking order changed: Not recorded
Supporting observations: Sep 7, 2026
Sources checked for latest review: Nov 5, 2026
Next category review due: Nov 7, 2026
Models covered by the published review: 5 of 11 eligible models

Still needing comparison: Rork + Current Model Stack; Superapp + Current Model Stack; Vibecode + Current Model Stack; VULK + Current Model Stack; a0.dev + Current Model Stack; Dreamflow + Current Model Stack.

The count includes models with evidence gaps. Rechecking sources does not make older tests new.

Small score gaps are close calls. Media rankings are editorial. How we compare →

The cost of a token

API pricing

Leading text & coding models. US dollars per million tokens.

Prices checked Sep 26, 2026

ModelInput 0 — $10Output 0 — $50

Claude Opus 5.5

Input$4

Output$20

Claude Opus 5

Input$5

Output$25

Claude Fable 5.1

Input$10

Output$50

GPT-6 Astra

Input$10

Output$50

Each column has its own scale. Standard uncached input and output rates; short-context pricing where applicable.

Input is what you send; output is what the model generates, including billable thinking. More thinking can mean more tokens at the same rate. These are API costs, not monthly subscriptions or a price per completed task.

Sources, thinking levels & pricing details

Compare standard paid API access. Caching, batch discounts, fast service, regional routing, tool fees and taxes can change the bill. Coding tools may add their own charges. Low, medium and high thinking are provider-specific settings, not comparable amounts of reasoning.

Gemini 3.8 Flash

Paid standard introductory rate through December 31, 2026; scheduled to become $1.50 input / $7.50 output on January 1, 2027. Output includes thinking tokens.

Provider pricing ↗Checked 2026-09-26

GPT-6 Sol

Standard rate for prompts up to 272K input tokens. Above 272K, the full request costs 2× the input rate and 1.5× the output rate. Low, medium and high reasoning share these token rates; extra reasoning adds billable output tokens.

Provider pricing ↗Checked 2026-09-26

Claude Opus 5.5

Global standard rate across the supported context window. Low, medium and high effort share these token rates; thinking adds billable output. Fast mode costs extra.

Provider pricing ↗Checked 2026-09-26

GPT-5.6 Sol

Promotional standard short-context rate, available at least through November 21, 2026. Above 272K input tokens, the full request costs 2× input and 1.5× output. Thinking adds billable output tokens.

Provider pricing ↗Checked 2026-09-26

Claude Opus 5

Global standard rate across the supported context window. Thinking is included in billed output; fast mode costs extra.

Provider pricing ↗Checked 2026-09-26

Claude Fable 5.1

Source checked Nov 5, 2026 · Next review target Nov 7, 2026

What changed

Past ranking and site updates

September 26, 2026

Opus 5.5 High enters the coding podium.

Today’s external-evidence review puts Claude Code + Opus 5.5 High first in Best Coding Overall and Code Generation, and second in Agentic Coding. These early recommendations carry an asterisk and a two-day review cycle. Exact effort settings, contrary results, and test limitations remain visible. Coding reviews proceed independently of pending writing approvals.

September 25, 2026

Clear recommendations. Reasons you can inspect.

Image and video reviews now support explicit first, second and third places, with category-specific criteria and published tiebreak decisions. Photorealism covers everyday phone snapshots as well as professional photography. Confidence and evidence gaps remain visible alongside our recommendations.

September 22, 2026

Image and video evidence, with its limits.

New source observations identify tested variants, task-specific findings, and conflicts with existing recommendations. They appear separately from the evidence behind the older rankings. Generic video summaries now receive the same provisional treatment as image summaries; scores and ranking-verification dates remain unchanged.

September 21, 2026

Clearer comparisons. Evidence you can inspect.

Larger text, direct category links, and expandable comparisons make the rankings easier to use. Evidence gaps are now visible, including inherited image placements awaiting version-specific confirmation. Ranking order and verification dates have not been advanced by this presentation update.

September 10, 2026

ChatGPT Images 2.5 replaces GPT Image 2.

OpenAI’s new image release adds faster generation, stronger subject preservation, more reliable multi-turn edits, and new direct-editing tools. The release initially inherited GPT Image 2’s placements. Those inherited placements are provisional until version-specific comparative evidence is documented. GPT-6 Astra’s access and API pricing have also been brought up to date.

September 8, 2026

Coding now ranks tested configurations.

Every Coding result identifies the system, underlying model or platform-selected AI, access plan, system version, and test date. Technical awards cleared the first supervised sweep; Builder awards are provisional until the full Practical Consensus threshold is met.

Methodology

Evidence in.
A useful answer out.

Task-specific evidence, clear recommendations, and honest limits. Explore our criteria, sources, and how close calls are decided.

Read the full methodology →

Best AI Models & Tools

Claude Code + Opus 5.5 High

Codex + GPT-6 Astra

Claude Code + Fable 5.1

Why doesn’t the overall winner win every category?

Codex + GPT-6 Astra

Claude Code + Opus 5.5 High

Claude Code + Fable 5.1

Claude Code + Opus 5.5 High

Codex + GPT-6 Astra

Claude Code + Fable 5.1

Claude Code + Fable 5.1

Codex + GPT-6 Astra

Claude Code + Opus 5

Cursor + Fable 5.1

Codex + GPT-6 Astra

Framer

Claude Code + Fable 5.1

Lovable

Bolt

Base44

FlutterFlow

Draftbit

Replit Agent

Know when the recommendations change.

API pricing

Gemini 3.8 Flash

GPT-6 Sol

Claude Opus 5.5

GPT-5.6 Sol

Claude Opus 5

Claude Fable 5.1

GPT-6 Astra

On the Horizon

GPT-6 Cyber

Gemini Omni 1.2 Flash

Tencent Hy4

GPT-Image-2.6 Sunburst and Flare

Canva Brush

Gemini 4 Carbon

Particle-1

Nemotron 3.5 Bolt E4B

Canva Craft

FLUX 3 Image Klein

Krea 3

GitHub Copilot HydraFusion / Max — October update

Muse Wonder Full

GLM-Legal

Mellum2.1 Thinking

GPT-6 Sol — October 2026 ChatGPT

MAI-Code-1.1-Flash

Kandinsky 6.0

Rho-1

Prism — research preview

Kolibri

Opus 5.5 High enters the coding podium.

Clear recommendations. Reasons you can inspect.

Image and video evidence, with its limits.

Clearer comparisons. Evidence you can inspect.

ChatGPT Images 2.5 replaces GPT Image 2.

Coding now ranks tested configurations.

Evidence in.A useful answer out.

Evidence in.
A useful answer out.