Atlas

Benchmarks

← All benchmarks

Toloka Arena (v1) pass^1

Agents

Per-trial success rate on Toloka Arena's superseded v1 task set, from the same five trials per task as its composite pass^5 headline: pass^1 is the share of single trials that succeed, the mean over the five trials.

Top models (higher is better)

ModelScore
Claude Fable 569.8
Claude Fable 5.168.2
Gemini 3.7 Flash66.9
Gemini 3.6 Flash66.1
Opus 4.865.7
GPT-5.564.6
GPT-6 Astra64.5
Claude Opus 564.4
Grok 4.564.1
GPT-5.463.6
Grok 4.663.5
Opus 4.762.1
GPT-5.6 Sol60.2
GLM-5.259.2
GLM-5.358.7
GPT-5.6 Terra57.7
Sonnet 557.5
Grok 4.757.3
Opus 4.656.4
Kimi K354.4
Gemini 3.5 Flash52.9
Muse Spark 1.152.8
Muse Spark 1.251.5
Gemini 3.1 Pro Preview51.4
Sonnet 4.650.2
GLM-5.148.5
Gemini 3 Flash Preview46.7
Kimi K2.637.4
MiniMax M334.0
Mistral Medium 3.532.2
gpt-oss-120b10.6
Grok 4.206.7
Nova 2 Lite3.7
Loading Atlas data…