Atlas

Benchmarks

← All benchmarks

LMArena Agent Arena Steering Burden

Agents

Signed percentage-point Agent Arena steering_burden diagnostic, displayed by the source as Steerability. Measures response to user corrections. Kept separate from the earlier steerability source signal because its key and observation population changed on the September 30 snapshot.

Top models (higher is better)

ModelScore
Claude Opus 5.517.6
Claude Fable 5.115.3
GPT-6 Astra14.5
Gemini 4 Argon14.4
GPT-6.1 Sol13.7
Muse Wonder Full11.8
GPT-6 Sol11.1
Claude Sonnet 5.510.8
Claude Fable 59.6
GPT-5.6 Sol8.4
Kimi K3.1 Pro8.3
Claude Opus 56.7
Sonnet 56.0
GPT-6 Luna5.7
Muse Spark 1.34.8
MiMo-V2.6-Flash4.7
Claude Haiku 5.54.5
Muse Wonder3.7
DeepSeek V4.1 Flash 10173.6
MiMo-V2.6-Pro2.1
GPT-5.6 Terra2.0
Hy4 Preview1.6
Qwen3.8-Max1.4
GLM-5.21.3
Gemini 3.8 Flash1.2
DeepSeek V4.1 Flash1.0
Grok 4.60.5
Step 5 Preview0.4
Kimi K30.3
Mistral Large 4 Preview0.1
GLM-5.3-0.2
DeepSeek V4 Pro 0813-0.4
Hy3-0.8
Grok 4.7-1.5
Qwen3.8-Flash-Next-2.1
GLM-5.3 Flash-3.3
Qwen3.8 27B-3.7
Gemini 3.7 Flash-4.5
Gemini 3.1 Pro Preview-6.4
Muse Spark 1.2-9.2
Loading Atlas data…