OPEN EVALUATION / OSWORLD-v2

Beyond the final score.

Follow the reasoning, code, and observations behind each result. Explore how eight frontier models get work done on a real desktop.

Read the research ↗

8frontier models

108tasks per model in the archive

3,480attempts, including retries

720pdesktop observations

THE LEADERBOARD

One harness. Eight perspectives.

Sort by

OpenGCUA full configuration on 82 offline OSWorld-v2 tasks
#	Model	Partial score ↑	Full success ↑	Steps / task ↓	Output tokens / task ↓	Explore
Loading evaluation results…

Leaderboard: the paper’s 82 offline tasks, full GCUA configuration (Table 2). The archive contains all 108 tasks per model and retains scored, unscored, empty, and retried attempts. Task scores below belong to the selected attempt; they are not the leaderboard aggregate.

A CLOSER LOOK

Choose a model to explore its tasks. Replay observations step by step, inspect the agent’s code and tool feedback, and compare the attempts behind a result.