OSWorld 2.0 (Meta, Sep 2026)Agents

The source chart does not identify the task release, scoring mode, or agent harness. This result is kept separate from first-attempt, partial, and strict OSWorld 2.0 variants.

Ranking Sources

#ModelScore

1

Claude Opus 5Anthropic · max mode

68.3%

2

Muse Spark 1.3Meta · max mode

66.9%

3

GPT-5.6 SolOpenAI · max mode

62.7%

4

Muse Spark 1.2Meta · xhigh mode

47.6%

4 models total

Back to benchmarks