← All benchmarks

OSWorld-Verified (361 tasks excluding Google Drive)

Agents

The 361-task OSWorld-Verified subset in Meta's Muse Spark Open evaluation excludes Google Drive tasks from the 369-task suite. The reported GUI-agent success rate averages four attempts per task with a 200-step limit. This explicit exclusion cohort remains distinct from the full suite and from 361-task cohorts whose exclusions have not been established.

Top models (higher is better)