Benchmark · Arenas: people's votes

LMArena Agent Arena

How models do as agents in sessions people run through the arena, scored with inverse propensity scoring (IPS) from signals in those sessions (task outcome, bash reliability, steerability, tool hallucination, praise and complaints) rather than from pairwise votes.

Score
IPS score, higher is better
Run by
LMArena
Status
Current, newest result 18 November 2026
Updated
daily
Data
the published file, CC-BY-4.0, read 20 November 2026
Method
arena.ai/blog/agent-arena-methodology/

Results

40 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Claude Fable 5.2 A0.2740.245 to 0.304Independenteffort max2026-11-18source
2Claude Opus 5.5 A0.1860.166 to 0.207Independenteffort max2026-11-18source
3Gemini 4 Neon B0.1840.159 to 0.209Independenteffort high2026-11-18source
4Gemini 4 Carbon B0.1610.154 to 0.168Independenteffort high2026-11-18source
5Claude Sonnet 5.5 A0.1560.127 to 0.185Independenteffort max2026-11-18source
6GPT-6.2 Astra AA0.1520.125 to 0.180Independenteffort max2026-11-18source
7GPT-6 Astra A0.1460.125 to 0.166Independenteffort max2026-11-18source
8Claude Fable 5.1 A0.1370.116 to 0.157Independenteffort max2026-11-18source
9Gemini 4 Hydrogen B0.1350.118 to 0.153Independenteffort high2026-11-18source
10GPT-6.1 Sol AA0.1130.102 to 0.124Independenteffort max2026-11-18source
10GPT-6.2 Sol AA0.1130.089 to 0.137Independenteffort max2026-11-18source
12GPT-6.1 Terra AA0.1050.093 to 0.117Independenteffort max2026-11-18source
13Gemini 4 Argon B0.1030.094 to 0.112Independenteffort high2026-11-18source
14GPT-6 Sol A0.0650.049 to 0.080Independenteffort max2026-11-18source
15Tencent Hy4 B0.0610.049 to 0.073Independent2026-11-18source
16Kimi K3.1 Pro D0.0390.028 to 0.050Independenteffort max2026-11-18source
17Gemini 3.8 Flash AA0.0150.006 to 0.024Independenteffort high2026-11-18source
17Grok 4.7 BB0.015-0.000 to 0.030Independenteffort xhigh2026-11-18source
17Inkling 2 Small C0.0150.005 to 0.025Independent2026-11-18source
20Muse Spark 1.3 D0.0110.002 to 0.020Independenteffort max2026-11-18source
21Muse Ember D0.0100.000 to 0.019Independenteffort max2026-11-18source
22DeepSeek V4.1 Flash C0.0080.003 to 0.013Independenteffort max2026-11-18source
23GPT-6 Luna A0.0060.002 to 0.011Independenteffort max2026-11-18source
23Muse Wonder D0.006-0.007 to 0.019Independenteffort max2026-11-18source
25GLM-5.3 B-0.000-0.009 to 0.008Independenteffort max2026-11-18source
26Kimi K3 B-0.001-0.008 to 0.005Independenteffort max2026-11-18source
27Claude Haiku 5.5 A-0.006-0.016 to 0.004Independenteffort high2026-11-18source
28MiniMax-M3.1-Flash C-0.037-0.051 to -0.022Independent2026-11-18source
29GLM-5.3-Flash B-0.039-0.046 to -0.031Independent2026-11-18source
30Qwen3.8-Flash-Next C-0.049-0.057 to -0.041Independent2026-11-18source
31Step 5 Preview E-0.051-0.066 to -0.037Independent2026-11-18source
32Qwen3.8-27B B-0.055-0.064 to -0.046Independent2026-11-18source
33Gemini 3.1 Pro (preview) AA-0.057-0.074 to -0.039Independent2026-11-18source
34Mistral Large 4 C-0.071-0.092 to -0.051Independent2026-11-18source
35Hy3 B-0.086-0.102 to -0.071Independent2026-11-18source
36MiniMax M3 C-0.109-0.121 to -0.097Independent2026-11-18source
37Mistral Medium 3.5 B-0.145-0.162 to -0.127Independent2026-11-18source
38Nemotron 3.5 Super VL C-0.173-0.182 to -0.164Independentprecision NVFP42026-11-18source
39Inkling B-0.185-0.200 to -0.169Independent2026-11-18source
39Inkling-Small B-0.185-0.206 to -0.165Independent2026-11-18source
Every setting, 46 rows
ModelAs the leaderboard names itScoreWho ran itSettingDate
Claude Fable 5.2Claude Fable 5.2 (Max)0.274Independenteffort max2026-11-18source
Claude Opus 5.5Claude Opus 5.5 (Max)0.186Independenteffort max2026-11-18source
Gemini 4 NeonGemini 4 Neon (High)0.184Independenteffort high2026-11-18source
Gemini 4 CarbonGemini 4 Carbon (High)0.161Independenteffort high2026-11-18source
Claude Sonnet 5.5Claude Sonnet 5.5 (Max)0.156Independenteffort max2026-11-18source
GPT-6.2 AstraGPT-6.2 Astra (Max)0.152Independenteffort max2026-11-18source
GPT-6 AstraGPT-6 Astra (Max)0.146Independenteffort max2026-11-18source
Claude Fable 5.1Claude Fable 5.1 (Max)0.137Independenteffort max2026-11-18source
Claude Opus 5.5Claude Opus 5.5 (High)0.137Independenteffort high2026-11-18source
Gemini 4 HydrogenGemini 4 Hydrogen (High)0.135Independenteffort high2026-11-18source
GPT-6.1 SolGPT-6.1 Sol (Max)0.113Independenteffort max2026-11-18source
GPT-6.2 SolGPT-6.2 Sol (Max)0.113Independenteffort max2026-11-18source
GPT-6.1 TerraGPT-6.1 Terra (Max)0.105Independenteffort max2026-11-18source
Gemini 4 ArgonGemini 4 Argon (High)0.103Independenteffort high2026-11-18source
GPT-6.1 SolGPT-6.1 Sol (Medium)0.082Independenteffort medium2026-11-18source
Claude Sonnet 5.5Claude Sonnet 5.5 (High)0.074Independenteffort high2026-11-18source
GPT-6 SolGPT-6 Sol (Max)0.065Independenteffort max2026-11-18source
Tencent Hy4Hy40.061Independent2026-11-18source
GPT-6.1 TerraGPT-6.1 Terra (Medium)0.046Independenteffort medium2026-11-18source
Kimi K3.1 ProKimi K3.1 Pro (Max)0.039Independenteffort max2026-11-18source
Gemini 3.8 FlashGemini 3.8 Flash (High)0.015Independenteffort high2026-11-18source
Grok 4.7Grok 4.7 (xHigh)0.015Independenteffort xhigh2026-11-18source
Inkling 2 SmallInkling 2 Small0.015Independent2026-11-18source
Muse Spark 1.3Muse Spark 1.3 (Max)0.011Independenteffort max2026-11-18source
Muse EmberMuse Ember (Max)0.010Independenteffort max2026-11-18source
DeepSeek V4.1 FlashDeepSeek-V4.1-Flash-20261017 (Max)0.008Independenteffort max2026-11-18source
GPT-6 LunaGPT-6 Luna (Max)0.006Independenteffort max2026-11-18source
Muse WonderMuse Wonder (Max)0.006Independenteffort max2026-11-18source
GLM-5.3GLM-5.3 (Max)-0.000Independenteffort max2026-11-18source
Kimi K3Kimi K3 (Max)-0.001Independenteffort max2026-11-18source
Claude Haiku 5.5Claude Haiku 5.5 (High)-0.006Independenteffort high2026-11-18source
GPT-6 LunaGPT-6 Luna (Medium)-0.031Independenteffort medium2026-11-18source
MiniMax-M3.1-FlashMiniMax-M3.1-Flash-0.037Independent2026-11-18source
GLM-5.3-FlashGLM-5.3-Flash-0.039Independent2026-11-18source
Qwen3.8-Flash-NextQwen3.8-Flash-Next-0.049Independent2026-11-18source
Step 5 PreviewStep 5 Preview-0.051Independent2026-11-18source
Qwen3.8-27BQwen3.8-27B-0.055Independent2026-11-18source
Gemini 3.1 Pro (preview)Gemini 3.1 Pro Preview-0.057Independent2026-11-18source
GPT-6 LunaGPT-6 Luna (Low)-0.064Independenteffort low2026-11-18source
Mistral Large 4Mistral Large 4-0.071Independent2026-11-18source
Hy3Hy3-0.086Independent2026-11-18source
MiniMax M3MiniMax-M3-0.109Independent2026-11-18source
Mistral Medium 3.5Mistral Medium 3.5-0.145Independent2026-11-18source
Nemotron 3.5 Super VLNVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4)-0.173Independentprecision NVFP42026-11-18source
InklingInkling-0.185Independent2026-11-18source
Inkling-SmallInkling-Small-0.185Independent2026-11-18source

A rank is among the models we list with a result, not a place on the whole leaderboard, and a model offered by several hosts is shown once, on its own listing. These results don't move our grades. Data: LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0, CC-BY-4.0.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.