← All benchmarks

OSWorld 2.0 - pre-August-8, 500 steps, maximum of three partial scores

Agents

OSWorld 2.0 partial reward reported as the maximum of three single-attempt runs, using the official Docker environment and evaluator at 1080p with a 500-step limit. Runs preceded the official August 8 patch. The exact dated task roster is not identified. This selected maximum is diagnostic and is not interchangeable with single-run or mean-run OSWorld partial scores.

Top models (higher is better)