OSWorld-G: leaderboard

A model sees a desktop screenshot with a plain-language instruction and must click inside the annotated target element, or refuse when the element is absent; the score is accuracy over all 564 samples with the original instructions. Unlike ScreenSpot it tags each sample by capability and includes fine-grained and infeasible targets; the refined-instruction version is OSWorld-G-R. The paper's authors ran every model; the Jedi models are their own, trained for GUI grounding.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Jedi-7B-1080p54.1
2Jedi 3B-1080p50.9
3Qwen 2.5 VL 32B Instruct46.5#1074
4UGround-V1-7B36.4
5Qwen 2.5 VL 7B Instruct31.4#1893
6Qwen 2.5 VL 3B Instruct27.3#2648

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=osworld-g · How It Works · Data refreshed daily, snapshot 2026-11-17.