OSWorld
Real-world computer use tasks in desktop OS environments
Models ranked
5
tracked on this benchmark
Score band
79.6 – 62.6
best → lowest tracked
Snapshot trend
+5.00
Oct 17 → Oct 26 · 1 models
Leaderboard
Tracked models ranked by Score (higher is better).
Notes: Moonshot self-reported launch table, row 'OSWorld V2.1', column 'Kimi K3.1 Pro (max)'. Settings: reasoning effort max, temperature 1.0, top-p 0.95 (blog footnote). Footnote: 'evaluated with the Kimi Agent harness in the official OSWorld v2 repo'.
Configuration: gpt-6.1-terra (Max)
Notes: OpenAI launch post chart 'OSWorld 2.1, offline set' (Max 0.742; Low 0.538, Medium 0.667, High 0.687, Xhigh 0.713). Offline set, harness and run count not stated. Retrieved 2026-10-26.
Configuration: Gemini 4 Argon; max of 3 runs
Notes: Blog table + methodology page: OSWorld 2.0 offline subset, partial score, max over 3 runs, Gemini CUA harness, official evaluator, 1080p, 500 steps, 08.08 patch. Comparability caveat: max-of-3 and offline subset differ from other OSWorld 2.0 rows. Retrieved 2026-10-17.
Notes: Official GPT-5.6 GA launch benchmark row.
How to read this benchmark
This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.
Trust this score when
- There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
- The model list covers the same version family you can actually deploy today.
- Top candidates overlap with your required routing and feature requirements.
Be cautious when
- There is only one benchmark snapshot or the dataset appears stale.
- The benchmark metric direction is opposite of your decision objective.
- The score difference between options is narrow and likely within implementation variance.
Related benchmarks
Last reviewed: May 20, 2026