OSWorld-Verified

369 computer-use tasks on a real Ubuntu desktop (browser, office apps, GIMP, VS Code), run and checked by the OSWorld team. Models only, not agent frameworks; one row per step budget. Near saturation.

CategoryAgents
MetricSuccess rate (%)
SourceOSWorld-Verified (original)
LicenseNone stated
Benchmark pageos-world.github.io
StatusScored
Weight0.6 within Agents
FitDiscrimination 0.98, noise 0.29
LineageShares weight with osworld

Results

#ModelProviderResultDate
1 Claude Fable 5 Anthropic 86.0% 2026-08-01
2 Claude Opus 5 Anthropic 83.4% 2026-08-01
3 Muse Spark 1.1 Meta 80.7% 2026-07-09
4 MiniMax-M3 MiniMax 75.2% 2026-06-08
5 Qwen3.7 Plus Alibaba 73.3% 2026-05-25
6 Kimi K2.6 Moonshot AI 73.1% 2026-04-20
7 Claude Sonnet 4.6 Anthropic 72.1% 2026-03-08
8 Kimi K2.5 Moonshot AI 63.3% 2026-01-31
9 Claude Sonnet 4.5 Anthropic 62.9% 2025-10-31
10 Seed 1.8 ByteDance 61.9% 2025-12-18
11 Claude Sonnet 4.5 Anthropic 58.1% 2025-10-31
12 Claude Sonnet 4 Anthropic 43.9% 2025-07-28
13 Claude Sonnet 4.5 Anthropic 42.9% 2025-10-31
14 Qwen3 VL Flash (Alibaba Cloud) Alibaba 41.6% 2025-10-23
15 Claude Sonnet 4 Anthropic 41.4% 2025-07-28
16 Claude Sonnet 3.7 Anthropic 35.8% 2025-07-28
17 Claude Sonnet 3.7 Anthropic 35.6% 2025-07-28
18 computer-use-preview-2025-03-11 Unknown 31.4% 2025-07-28
19 computer-use-preview-2025-03-11 Unknown 31.3% 2025-07-28
20 Claude Sonnet 4 Anthropic 31.2% 2025-07-28
21 Claude Sonnet 3.7 Anthropic 27.1% 2025-07-28
22 computer-use-preview-2025-03-11 Unknown 26.0% 2025-07-28
23 o3 OpenAI 23.0% 2025-07-28
24 o3 OpenAI 17.2% 2025-07-28
25 o3 OpenAI 9.1% 2025-07-28
26 Qwen2.5-VL 72B Instruct Alibaba 5.0% 2025-07-28
27 Qwen2.5-VL 72B Instruct Alibaba 4.4% 2025-07-28