Benchmark

OSWorld Benchmark

What is OSWorld?

Screen resolution, OS image and action space all change results. The Verified and 2.0 revisions are not comparable to the original.

Released: 2024-04-11
Evaluates: computer use
Openness: unknown
Importer: Model reports
Reported scores: 4

Source provenance

Original evidence https://os-world.github.io/

OSWorld links

Website https://os-world.github.io/

OSWorld results and reported scores

Model reports

Model	Organization	Reported value	Reported	Evidence
Claude Opus 5	Anthropic	70.57	2026-07-24	evidence
Kimi K3	Moonshot AI	58.3	2026-06-13	evidence
Kimi K3	Moonshot AI	84.8	2026-06-13	evidence
Qwen3.5-397B-A17B	Qwen	62.2	2026-02-16	evidence

Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.

Which sources cite OSWorld?

Qwen3.8-27B Qwen, 2026-08-05
Claude Opus 5 Anthropic, 2026-07-24
Kimi K3 Moonshot AI, 2026-06-13
Claude Fable 5 and Mythos 5 Anthropic, 2026-06-09
Qwen3.5-397B-A17B Qwen, 2026-02-16