Tianbao Xie et al.

OSWorld-Verified

OSWorld-Verified is the July 2025 repaired release of OSWorld's real-computer evaluation. It measures whether a model-agent system can finish desktop and web tasks from configured starting states, with success checked by execution-based evaluators.

In index
Agentic
Source ↗

Models34

Top result86.1%Qwen3.8 Max

Top-3 spread1.1 pts

Year2025

Result and price

Results

34 results

#ModelResult

Qwen3.8 Max86.1%

Claude Fable 585.0%

Claude Mythos 585.0%

Qwen3.8-27B84.3%

Claude Opus 4.883.4%

Gemini 3.6 Flash83.0%

07HCHolo3-35B-A3B82.6%

MiMo-V2.6-Pro82.0%

Claude Sonnet 581.2%

Muse Spark 1.180.8%

MiMo-V2.6-Flash80.8%

12HCHolo3-122B-A10B78.8%

GPT-5.578.7%

Gemini 3.5 Flash78.4%

Claude Opus 4.7 (Adaptive)78.0%

16TEUI-Mate-27B77.0%

GPT-5.475.0%

Gemini 3.5 Flash-Lite74.0%

Qwen3.7 Plus73.3%

20MAKimi K2.673.1%

Claude Opus 4.672.7%

Claude Sonnet 4.672.1%

GPT-5.4 mini72.1%

MiniMax M370.1%

Claude Opus 4.566.3%

26TEUI-Mate-9B66.2%

Muse Glimmer 30B65.9%

GPT-5.3 Codex64.7%

Claude Sonnet 4.561.4%

Qwen3.5-122B-A10B58.0%

Qwen3.5-27B56.2%

Qwen3.5-35B-A3B54.5%

GPT-5.247.3%

GPT-5.4 nano39.0%