Best models overall

LiveBench and MathArena combined · 8 models with both · updated 2026-09-24

Top combined index

92.9Claude Fable 5.1 & GPT-6 Astra

Tied by different routes. Claude Fable 5.1: 100th LiveBench / 86th MathArena; GPT-6 Astra: 86th LiveBench / 100th MathArena percentile.

Most uneven

43 ptsQwen3.8 2.4T A95B

29th percentile on LiveBench vs 71th on MathArena. The average hides this gap.

Best open weights

50.0Qwen3.8 2.4T A95B

Ranks #4 of 8 on the combined index.

What makes up the index

Each bar is half LiveBench percentile, half MathArena percentile, among the models that have both.

LiveBench shareMathArena share
020406080100Claude Fable 5.1LiveBench share: 50.0MathArena share: 42.992.9GPT-6 AstraLiveBench share: 42.9MathArena share: 50.092.9Muse Spark 1.3LiveBench share: 35.7MathArena share: 21.457.1Qwen3.8 2.4T A95BLiveBench share: 14.3MathArena share: 35.750.0DeepSeek V4.1 FlashLiveBench share: 28.6MathArena share: 14.342.9Grok 4.7LiveBench share: 7.1MathArena share: 28.635.7Kimi K3LiveBench share: 21.4MathArena share: 0.021.4Gemini 3.8 FlashLiveBench share: 0.0MathArena share: 7.17.1

All results

RankModelCombined indexLiveBenchMathArena
1=Claude Fable 5.1Anthropic
92.9
83.4claude-fable-5-1-max-effort87.72% ± 6.03post-releaseClaude-Fable-5.1 (high)
1=GPT-6 AstraOpenAI
92.9
82.2gpt-6-astra-max88.60% ± 5.83post-releaseGPT-6 Astra (max)
3Muse Spark 1.3Meta
57.1
81.6muse-spark-1.3-xhigh45.61% ± 9.91post-releaseMuse Spark 1.3
4Qwen3.8 2.4T A95BQwen · Open weights
50.0
78.5qwen3.8-max57.89% ± 12.82post-releaseQwen3.8-Max
5DeepSeek V4.1 FlashDeepSeek · Open weights
42.9
81.1deepseek-v4.1-flash-max43.86% ± 9.11post-releaseDeepSeek-V4.1-Flash (Max)
6Grok 4.7SpaceXAI
35.7
77.4grok-4.7-xhigh46.49% ± 9.78post-releaseGrok 4.7 (xhigh)
7Kimi K3MoonshotAI · Open weights
21.4
79.2kimi-k338.60% ± 12.64Kimi K3 (Think)
8Gemini 3.8 FlashGoogle
7.1
75.8gemini-3.8-flash-high40.35% ± 9.05post-releaseGemini 3.8 Flash
51 more with only one source — not ranked

How the ranking works

Combined index
Each model’s LiveBench and MathArena scores are turned into a percentile among the 8 models that have both, then averaged 50/50. Ties use average ranks. It shows relative order, not accuracy.
Who gets ranked
Only models with results from both sources. Missing results are never filled with zero. Search and filters don’t change who’s compared.
Read with care
Math counts for half, so this isn’t a coding ranking. Use By task for that. Reasoning settings and test dates differ between models. Small gaps aren’t decisive.
post-release
MathArena flags models released after its test questions were public. A contamination risk, not proof.

LiveBenchGeneral capability · edition 2026-06-25 · 63 configurations · checked 2026-09-24 · Published results ↗

MathArenaArXivMath · August 2026 · 10 configurations · checked 2026-09-24 · Published results ↗