OSWorld-Verified
Computer-use tasks
agents and tool usecrowded
Definition
- Version
- not in the notes
- Category
- agents and tool use
- Maintainer
- XLANG Lab (HKU)
- Status
- crowded
- Home page
- https://arxiv.org/abs/2404.07972
Sources
- openai.com/index/introducing-gpt-5-3-codex https://openai.com/index/introducing-gpt-5-3-codex
- huggingface.co/moonshotai/Kimi-K3 https://huggingface.co/moonshotai/Kimi-K3
- OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments arxiv.org https://arxiv.org/abs/2404.07972
- OSWorld-Verified Leaderboard & Scores , October 2026 benchlm.ai https://www.benchlm.ai/benchmarks/osworld-verified
Scores at a Glance
Score in %
Higher is better for this unit.
0255075100
Where this comes from
Each bar is one model, with its most independent recorded score. Hatched bars are vendor-reported; the word beside each bar says who ran it. Settings and dates differ between rows; the Scores table below has every result.
Scores
| Model | Score | Who ran it | Setting | View or population | Date | Source |
|---|---|---|---|---|---|---|
 Qwen3.8 Max | 86.1 % | vendor-reported | The genealogy notes label the holder open weights; not supported by its sources | rank 1, BenchLM aggregator, top score | 10 Oct 2026 | benchlm.ai |
 Kimi K3 | 84.8 % | vendor-reported | – | – | – | huggingface.co |
 GPT-5.3-Codex | 64.7 % | vendor-reported | xhigh | – | 5 Feb 2026 | openai.com |
Where this comes from
Scores from different settings, harnesses or dates are not directly comparable; the row says which. Sorted by score, highest first, as a reading aid, not a ranking.
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "OSWorld-Verified", https://aurion.semai.id/benchmarks/osworld-verified/, records of October 2026.