Published tests · Sources linked

Performance.

Pick the work. See the evidence.

Work with an AI agent

Arena estimates a model’s contribution to agent outcomes using randomized model selection and feedback from real sessions.

Agent Arena ↗
Ratings published Oct 28, 2026

Original source checked. New permitted results are imported automatically. Last successful check: Oct 30, 2026, 7:41 AM UTC.

ModelsSearchLocal model sizeOrder

Maker and dates

MakerRelease / listing fromRelease / listing to

Dates are reported model releases or repository listings, not test dates. Date filters exclude unknown dates. Size uses the published model parameter count, including inactive experts. Extra vision or draft modules may add weights; each model’s size note explains what is included. This is not a memory estimate.

57 tested configurations

A net improvement index relative to Arena’s baseline. Zero is the baseline, not zero capability; this is not a task completion percentage. Each configuration stays separate. How this test works ↗

Model & tested setup	Net improvement indexHigher is better · 95% intervalChart scale: -0.230 to +0.240	Evidence	Access & size
Claude Opus 5.5Anthropic Claude Opus 5.5 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 22, 2026 · Release date reported by Models.dev; not independently verified	+0.209+0.188 to +0.231	13,584 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Gemini 4 Carbon (High)Google Gemini 4 Carbon (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.160+0.150 to +0.170	26,511 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Claude Sonnet 5.5Anthropic Claude Sonnet 5.5 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 28, 2026 · Release date reported by Models.dev; not independently verified	+0.138+0.112 to +0.165	8,300 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Claude Fable 5.1Anthropic Claude Fable 5.1 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 1, 2026 · Release date reported by Models.dev; not independently verified	+0.134+0.115 to +0.152	17,347 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-6 AstraOpenAI GPT-6 Astra (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 4, 2026 · Release date reported by Models.dev; not independently verified	+0.130+0.109 to +0.150	15,070 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-6.1 SolOpenAI GPT-6.1 Sol (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 29, 2026 · Release date reported by Models.dev; not independently verified	+0.125+0.114 to +0.135	28,883 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Gemini 4 ArgonGoogle Gemini 4 Argon (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Oct 13, 2026 · Release date reported by Models.dev; not independently verified	+0.107+0.099 to +0.115	50,974 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-6.1 Terra (Max)OpenAI GPT-6.1 Terra (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.106+0.086 to +0.125	8,977 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Muse Wonder Full (Max)Meta Muse Wonder Full (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.101+0.091 to +0.110	26,753 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Claude Fable 5Anthropic Claude Fable 5 (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Jun 7, 2026 · Release date reported by Models.dev; not independently verified	+0.080+0.069 to +0.091	50,866 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-6 SolOpenAI GPT-6 Sol (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 22, 2026 · Release date reported by Models.dev; not independently verified	+0.080+0.063 to +0.096	12,107 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Claude Opus 5Anthropic Claude Opus 5 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Jul 24, 2026 · Release date reported by Models.dev; not independently verified	+0.076+0.063 to +0.088	31,777 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Kimi K3.1 Pro (Max)Moonshot AI Kimi K3.1 Pro (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.060+0.044 to +0.075	14,300 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-5.6 SolOpenAI GPT-5.6 Sol (xHigh) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Jul 9, 2026 · Release date reported by Models.dev; not independently verified	+0.048+0.038 to +0.058	45,156 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Muse WonderMeta Muse Wonder (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Oct 12, 2026 · Release date reported by Models.dev; not independently verified	+0.031+0.017 to +0.046	12,530 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Claude Sonnet 5Anthropic Claude Sonnet 5 (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Jun 29, 2026 · Release date reported by Models.dev; not independently verified	+0.029+0.018 to +0.041	41,456 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
MiMo-V2.6-ProXiaomi MiMo-V2.6-Pro Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.028+0.019 to +0.036	46,354 sessionsRatings published Oct 28, 2026Original result ↗	Weights not verifiedSource-listed model license: MIT
DeepSeek-V4.1-Flash-20261017 (Max)DeepSeek DeepSeek-V4.1-Flash-20261017 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.027+0.020 to +0.033	54,255 sessionsRatings published Oct 28, 2026Original result ↗	Weights not verifiedSource-listed model license: MIT
Gemini 3.8 FlashGoogle Gemini 3.8 Flash (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 2, 2026 · Release date reported by Models.dev; not independently verified	+0.026+0.018 to +0.035	52,730 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Muse Spark 1.3Meta Muse Spark 1.3 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 2, 2026 · Release date reported by Models.dev; not independently verified	+0.026+0.018 to +0.034	80,518 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
Grok 4.7xAI Grok 4.7 (xHigh) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 21, 2026 · Release date reported by Models.dev; not independently verified	+0.024+0.009 to +0.039	18,559 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GPT-6 LunaOpenAI GPT-6 Luna (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Sep 22, 2026 · Release date reported by Models.dev; not independently verified	+0.022+0.018 to +0.026	217,544 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
DeepSeek-V4.1-Flash-20260910 (Max)DeepSeek DeepSeek-V4.1-Flash-20260910 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.019+0.016 to +0.022	266,436 sessionsRatings published Oct 28, 2026Original result ↗	Weights not verifiedSource-listed model license: MIT
Claude Haiku 5.5Anthropic Claude Haiku 5.5 (High) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Oct 7, 2026 · Release date reported by Models.dev; not independently verified	+0.014+0.004 to +0.025	29,343 sessionsRatings published Oct 28, 2026Original result ↗	Cloud · closed modelSource-listed model license: Proprietary
GLM-5.2 (Max)Z.ai GLM-5.2 (Max) Test settings & model date Exact orchestrator configuration is shown above. Evaluated in Arena Agent Mode; other harness components are not specified in this summary. Model release date not verified for this exact source label.	+0.013+0.005 to +0.020	98,542 sessionsRatings published Oct 28, 2026Original result ↗	Weights not verifiedSource-listed model license: MIT

Coverage, permissions & freshness

Overall orchestrator results only. This test uses a different scale from Code Arena and Terminal-Bench.

2 configurations have verified downloadable weights. 20 have no verified weight mapping; they are excluded from the local and cloud-only filters.

Attribution: Arena / lmarena-ai, leaderboard-dataset. CC BY 4.0 · source license ↗. Adapted by Privaj Scout: selected summaries, formatting, exact identity mapping, and filters. Source scores are unchanged. No endorsement implied.

Our server checks the original sources every six hours. Downloads are cached and new results are checked before they appear. Failed checks keep the last verified data and retry automatically. Publication dates above describe the imported results, not when we downloaded them.

Different tests are never averaged into an overall score. Human preference, agent outcomes, and terminal task completion answer different questions.