Atlas

Benchmarks

← All benchmarks

Toloka Arena (v3) pass^1

Agents

Per-trial success rate on Toloka Arena's current (v3) task set, from the same five trials per task as the 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 Opus 5.584.4
Claude Fable 5.182.6
Gemini 4 Argon80.5
Claude Sonnet 5.579.8
GPT-6.1 Sol79.4
GPT-6 Astra78.4
Muse Wonder78.4
GPT-6 Sol77.4
GPT-5.6 Sol76.6
Grok 4.676.4
Claude Opus 576.0
Gemini 3.8 Flash75.6
Gemini 3.7 Flash72.6
GPT-6.1 Terra72.0
Beam 1.0 501B A23B64.2
GPT-6 Luna62.6
Qwen3.8 27B58.9
Grok 4.757.4
MiniMax M3.1 Flash51.1
Claude Haiku 5.549.6
Sonnet 545.0
Kimi K341.0
Muse Spark Open39.6
Nemotron 3.5 Super VL 120B A12B (precision unspecified)25.8
Gemini 3.1 Pro Preview22.4
Loading Atlas data…