Atlas

Benchmarks

← All benchmarks

Odysseys2 Perfect-Task Rate

Agents · 2026-10-16

Share of Odysseys2's 100 live-web browser-agent tasks on which an agent's attempt earns a perfect final score: every rubric item met and no reward-hacking flag, with no partial credit. It is the strict companion of the family's primary mean final score, computed from the same graded attempts.

Top models (higher is better)

ModelScore
GPT-6 Astra47.0
GPT-5.6 Luna22.0
Claude Opus 5.518.0
GPT-5.6 Sol15.0
Grok 4.713.0
Claude Fable 54.1
Claude Opus 54.0
Grok 4.54.0
Sonnet 52.0
GPT-5.6 Terra2.0
Gemini 3.1 Pro Preview0.0
Gemini 3.7 Flash0.0
Gemini 3.8 Flash0.0
GPT-5.4 Mini0.0
Loading Atlas data…