Benchmarks / Arena (formerly LMArena)

Reported by Arena (formerly LMArena)

Arena (formerly LMArena)

Estimated effect on the agent taking a user's correction.

Results dated
28 Sep 2026
Models
46
Unit
IPS effect estimate
Licence
Creative Commons Attribution 4.0 International
Arena Agent: steerability, IPS effect estimate, higher is better
#ModelArena Agent: steerability · 95% range
IPS effect estimate, higher is better
1 GPT-6 SolOpenAI
0.15
0.10–0.19
1 Claude Opus 5.5Anthropic
0.14
0.10–0.18
1 Claude Fable 5Anthropic
0.11
0.09–0.14
1 Claude Opus 5Anthropic
0.11
0.08–0.13
1 Claude Opus 4.8Anthropic
0.10
0.07–0.12
1 Claude Fable 5.1Anthropic
0.08
0.05–0.12
1 Claude Opus 5Anthropic
0.08
0.05–0.11
1 Claude Sonnet 5Anthropic
0.07
0.04–0.11
4 GPT-5.5OpenAI
0.07
0.05–0.09
4 GPT-5.6 SolOpenAI
0.07
0.04–0.09
6 Gemini 3.8 FlashGoogle
0.04
0.02–0.06
6 GLM 5.2Z.ai
0.04
0.03–0.06
6 GPT-5.6 TerraOpenAI
0.04
0.02–0.06
6 Grok 4.5xAI
0.04
0.02–0.06
6 Grok 4.6xAI
0.03
0.01–0.05
6 DeepSeek V4 Pro 0813DeepSeek
0.02
-0.02–0.06
6 GPT-6 LunaOpenAI
0.02
-0.02–0.06
10 Grok 4.7xAI
0.00
-0.03–0.04
11 GLM 5.3Z.ai
0.03
0.02–0.04
11 GPT-5.4OpenAI
0.02
0.01–0.04
11 GPT-5.5OpenAI
0.02
0.01–0.04
11 Qwen3.8 Max (0902)Alibaba
0.02
0.01–0.03
11 GPT-5.6 LunaOpenAI
0.02
0.01–0.03
11 Muse Spark 1.3Meta
0.02
0.00–0.03
11 GPT-6 AstraOpenAI
-0.01
-0.06–0.04
16 Gemini 3.7 FlashGoogle
-0.00
-0.01–0.01
17 Hy4 previewTencent
-0.01
-0.03–0.01
17 Hy3Tencent
-0.01
-0.04–0.01
20 Kimi K3Moonshot AI
-0.01
-0.02–0.01
21 Qwen3.8 27BAlibaba
-0.01
-0.02–0.00
22 GLM 5.3 FlashZ.ai
-0.01
-0.02–-0.00
22 DeepSeek V4 Pro 0423DeepSeek
-0.03
-0.05–-0.01
24 DeepSeek V4.1 FlashDeepSeek
-0.03
-0.05–-0.02
24 Qwen3.8 Flash NextAlibaba
-0.03
-0.05–-0.02
24 Qwen3.7 MaxAlibaba
-0.03
-0.05–-0.02
24 Gemini 3.6 FlashGoogle
-0.04
-0.06–-0.02
29 Gemini 3.1 Pro PreviewGoogle
-0.04
-0.05–-0.03
29 MiMo-V2.5-ProXiaomi
-0.05
-0.06–-0.03
31 Muse Spark 1.2Meta
-0.05
-0.07–-0.04
31 Mistral Medium 3.5Mistral
-0.07
-0.10–-0.04
33 Muse Spark 1.1Meta
-0.06
-0.07–-0.05
34 Qwen3.7 PlusAlibaba
-0.07
-0.10–-0.05
35 MiniMax M3MiniMax
-0.07
-0.09–-0.05
40 Solar Pro 4Upstage
-0.10
-0.14–-0.07
41 Inkling SmallThinkingmachines
-0.12
-0.15–-0.08
42 InklingThinkingmachines
-0.12
-0.14–-0.10

Models share a rank when their ranges overlap. Results as published by Arena (formerly LMArena); we do not re-run them.

What it measures

Estimated effect on the agent taking a user's correction.

What it does not measure

Not a success rate: the estimated effect of choosing this model as the agent's orchestrator, against an even mix of the models tested, inside Arena's own harness.

Source

Arena (formerly LMArena) · Methodology

Contains data from the Arena Leaderboard Dataset by Arena, licensed under CC BY 4.0. Licence: Creative Commons Attribution 4.0 International.