Atlas

Benchmarks

← All benchmarks

OSWorld 2 (Aug 8) offline - reported partial reward, step limit unspecified

Agents · 2026-08-08

Partial reward on the 82-task offline subset of the OSWorld 2.0 v2026.08.08 release, as reported by model developers who do not state the step limit. Sibling of the 500-step partial-score row; repeated-run averages do not license an item-level binomial noise model. Higher is better.

Top models (higher is better)

ModelScore
GPT-6 Astra72.6
GPT-6.1 Sol71.4
Claude Opus 570.2
GPT-5.6 Sol65.8
GPT-6 Sol64.4
Loading Atlas data…