OSWorld 2.0

Measures: An agent operates a real desktop (web apps and professional desktop software) to finish 108 long-horizon workflows that take humans a median of about 1.6 hours; scored as binary task completion at a 500-step budget, with a partial-credit checkpoint score reported alongside.

Maintainer: XLANG Lab, University of Hong Kong (Snorkel AI as research and data partner)
Introduced: 2026
Status: Active
Saturation: Headroom
Typical provenance: Mixed
In scoring set: Yes

Paper Official leaderboard

Leaderboard

Show independent results only

Every result we hold, whoever ran it. Only independent and third-party rows count towards a score.

Leaderboard
Model	Binary completion (500 steps)	Output tokens	Partial checkpoint score	Run cost (full task set)	Provenance	Source	Evaluated
Qwen3.7-Plus Alibaba Group	2.8%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Kimi K2.6 Moonshot AI	4.6%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
MiniMax-M3 MiniMax	4.6%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Sonnet 4.6 Anthropic · Max effort	8.3%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Sonnet 4.6 Anthropic · Medium effort	9.3%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
GPT-5.5 OpenAI · Xhigh effort	13.0%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 4.7 Anthropic · Max effort	13.9%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 4.7 Anthropic · Max effort	18.2%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 4.8 Anthropic · Max effort	18.5%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 4.8 Anthropic · Max effort	20.6%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 5 Anthropic · Low effort	22.3%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 5 Anthropic · Medium effort	25.3%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
GPT-5.6 Sol OpenAI · Max effort	27.3%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 5 Anthropic · High effort	29.0%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 5 Anthropic · Xhigh effort	30.2%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—
Claude Opus 5 Anthropic · Max effort	31.4%	—	—	—	Third party	https://osworld-v2.xlang.ai/ via Epoch AI	—

Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0.

Sources

Epoch AI