Atlas

Benchmarks

OSWorld 2.0

OSWorld 2.0 evaluates computer-use agents on 108 long-horizon real-world workflows spanning everyday and professional desktop/web tasks. The official leaderboard reports binary completion accuracy and partial score under fixed step budgets.

Models
0
Top score
—
Leader
—
Updated
—
Resolution
—
Saturation
—

No signal or difficulty estimate. No usable scores.

This is a benchmark-family page. Scores are recorded on the component benchmarks below rather than directly on the parent, because the components use different score scales and directions.

MetricModelsLeaderDirectionDate
OSWorld 2.0 500-step Partial Score14Claude Opus 570.6higher is better2026-07-09
OSWorld 2.0 500-step Binary Accuracy9Opus 4.820.6higher is better2026-07-09
OSWorld 2.0 300-step Binary Accuracy6Opus 4.713.0higher is better2026-10-15
OSWorld 2.0 300-step Partial Score6GPT-5.549.5higher is better2026-10-15
OSWorld 2.0 150-step Binary Accuracy5GPT-5.513.0higher is better2026-10-15
OSWorld 2.0 150-step Partial Score5GPT-5.546.7higher is better2026-10-15
OSWorld 2.0 - pre-August-8, 500 steps, maximum of three partial scores3Gemini 3.8 Flash59.0higher is better2026-08-13
OSWorld 2 (Aug 8) Full, 500-step Binary Accuracy2Claude Opus 531.4higher is better2026-10-15
OSWorld 2 (Aug 8) Offline, 500-step Binary Accuracy2Claude Opus 534.7higher is better2026-10-15
OSWorld 2.1 Offline, 500-step Binary Accuracy2Claude Fable 5.254.6higher is better2026-10-15
OSWorld 2 August 2026 with Anthropic fixes, 500-step binary1Claude Fable 5.141.7higher is better—
OSWorld 2.1 Full, 500-step Binary Accuracy1Claude Opus 544.3higher is better2026-10-15
Loading Atlas data…