← All benchmarks

OSWorld 2 August 2026 with Anthropic fixes, 500-step partial

Agents

Mean checkpoint partial credit on the full 108-task August 2026 OSWorld 2 release with subsequent author task fixes and Anthropic-reported setup and grading fixes. Uses 500 actions, 1080p screenshots and five independent runs per task; kept diagnostic because the exact unpublished patch set differs from the official released cohorts.

Top models (higher is better)