Benchmarks / Agentic

Agent Arena

Arena.ai Agent Arena: causal 'net improvement' a model contributes to real-world agentic sessions (task success, praise vs complaint, steerability, bash recovery, tool hallucination), estimated from millions of live Agent Mode sessions. Percent with 95% CI; not an Elo.

50models
–Open Index weight
–last updated
LMArenapublisher

Agent Arena

Higher is better · best effort per model

Chart settings
50 of 50 selected

full table below

#ModelEffortScoreCost / taskHarness
1Opus 5.5max17.2%––
2GPT-6.1 Solmax14.8%––
3Fable 5.1max14.1%––
4Gemini 4high13.7%––
5GPT-6 Astramax13.7%––
6Sonnet 5.5max12.5%––
7GPT-6 Solmax11.2%––
8Opus 5max9.2%––
9Fable 5high8.4%––
10GPT-5.6 Solxhigh7.4%––
11Opus 4.8high7.1%––
12Sonnet 5high4.7%––
13Gemini 3.8 Flashhigh4.0%––
14MiMo V2.6 Prohigh3.8%––
15Muse Spark 1.3max3.6%––
16Grok 4.7xhigh3.4%––
17GPT-5.5xhigh3.3%––
18GPT-6 Lunamax3.3%––
19DeepSeek V4.1 Flashmax3.0%––
20Hy4–2.9%––
21MiMo V2.6 Flashhigh2.5%––
22GLM 5.2max2.4%––
23GLM 5.3max2.0%––
24Kimi K3max1.9%––
25Qwen 3.8 Maxmax1.9%––
26Haiku 5.5high1.9%––
27GPT-5.4high1.6%––
28Grok 4.6xhigh1.2%––
29GPT-5.6 Terraxhigh0.4%––
30DeepSeek V4 Prohigh0.4%––
31Grok 4.5max0.1%––
32Gemini 3.7 Flashhigh-0.4%––
33GLM 5.3 Flashmax-1.1%––
34GPT-5.6 Lunaxhigh-1.8%––
35Qwen 3.8 Flashmax-2.6%––
36Mistral Large 4high-3.2%––
37Qwen 3.8 27Bmax-3.2%––
38Gemini 3.1 Prohigh-3.5%––
39Gemini 3.6 Flashhigh-4.7%––
40Qwen 3.7 Maxmax-5.8%––
41Hy3xhigh-6.1%––
42Muse Spark 1.1max-6.2%––
43Qwen 3.7 Plushigh-6.3%––
44Muse Spark 1.2xhigh-6.5%––
45MiMo V2.5 Proxhigh-7.7%––
46MiniMax M3max-7.9%––
47Nemotron 3 Ultra 550B A55Bmax-8.3%––
48Mistral Medium 3.5high-11.3%––
49Inklingmax-13.6%––
50Inkling Smallxhigh-14.7%––