Atlas

Benchmarks

← All benchmarks

LMArena Agent Arena Chat Tool Hallucination

Agents · 2026-08-14

Signed percentage-point Agent Arena Chat diagnostic improvement signal for reductions in apparent nonexistent or inappropriate tool use, recomputed on sessions tagged as general agentic chat such as creative writing, learning, everyday research and media generation. Atlas stores the raw improvement orientation relative to the category baseline: positive values indicate more grounded tool use and higher is better. Reported confidence intervals are not standard errors.

Top models (higher is better)

ModelScore
GPT-5.50.5
Claude Fable 50.5
GPT-6 Astra0.5
GLM-5.30.5
Claude Sonnet 5.50.5
Claude Opus 50.5
GPT-5.6 Terra0.5
Kimi K30.5
GPT-6.1 Sol0.5
Claude Fable 5.10.4
Muse Spark 1.30.4
Muse Wonder0.4
Gemini 3.1 Pro Preview0.4
Step 5 Preview0.4
Sonnet 50.4
GPT-5.6 Luna0.4
Grok 4.70.4
Claude Opus 5.50.4
GLM-5.20.3
GPT-5.40.3
GLM-5.3 Flash0.3
GPT-6 Luna0.3
GPT-6 Sol0.3
DeepSeek V4.1 Flash0.3
Muse Wonder Full0.3
Gemini 3.7 Flash0.3
Grok 4.60.3
DeepSeek V4.1 Flash 10170.2
Hy4 Preview0.2
Gemini 3.8 Flash0.2
Gemini 4 Argon0.1
GPT-5.6 Sol0.1
DeepSeek V4 Pro 08130.1
Muse Spark 1.10.0
Claude Haiku 5.50.0
Qwen3.7-Max (source unspecified)-0.1
MiniMax M3-0.1
Muse Spark 1.2-0.1
Qwen3.7-Plus (source unspecified)-0.3
Qwen3.8-Flash-Next-0.4
Loading Atlas data…