benchmarks.site

Public leaderboard · Agents, tool use & computer use

OSWorld 2.0

A much harder follow-up to OSWorld: the AI must carry out long, multi-step office and professional workflows on a real computer, jobs that take a person about an hour and a half. It shows whether AI can handle long, realistic computer work rather than quick clicks.

Leaderboard

Top 3 right now

  1. 1stClaude Opus 544.33
  2. 2ndMuseSpark 1.339.25
  3. 3rdClaude Opus 536.89

What it measures

108 long-horizon computer-use workflows across seven professional domains, scored by binary completion within a 500-step limit plus a partial checkpoint score.

How to read the numbers

Share of the 108 long workflows completed fully and correctly within 500 steps (binary accuracy, %) — higher is better.

Results

24 rows · higher is better (Binary accuracy (%))CSV
1Claude Opus 5Anthropicmaxbatch toolv2.144.3377.67yes
2MuseSpark 1.3MuseSparkmaxbatch toolv2.139.2569.56yes
3Claude Opus 5Anthropichighbatch toolv2.136.8968.92yes
4Claude Opus 5Anthropicxhighbatch toolv2.133.3373.11yes
5Claude Opus 5Anthropicmediumbatch toolv2.133.0164.85yes
6Claude Opus 5Anthropicmaxbatch toolv2026.08.0831.4368.31yes
7Claude Opus 5Anthropicxhighbatch toolv2026.08.0830.2167.67yes
8Claude Opus 5Anthropichighbatch toolv2026.08.0828.9863.79yes
9GPT-5.6 SolOpenAImaxbatch toolv2026.08.0827.3462.72yes
10MiniMax M3.1 FlashMiniMaxadaptivebatch toolv2026.08.0826.8562.01no
11Claude Opus 5Anthropicmediumbatch toolv2026.08.0825.3360.49yes
12Claude Opus 5Anthropiclowbatch toolv2026.08.0822.3254.89yes
13Claude Opus 4.8Anthropicmaxbatched toolv2026.06.2420.654.8yes
14Claude Opus 5Anthropiclowbatch toolv2.118.8159.09yes
15Claude Opus 4.8Anthropicmaxstandardv2026.06.2418.5249.33yes
16Claude Opus 4.7Anthropicmaxbatched toolv2026.06.2418.248.91yes
17MuseSpark 1.2MuseSparkxhighbatch toolv2.115.8444.08no
18Claude Opus 4.7Anthropicmaxstandardv2026.06.2413.949.1yes
19GPT-5.5OpenAIxhighbatch toolv2026.06.2413.049.5yes
20Claude Sonnet 4.6Anthropicmediumstandardv2026.06.249.333.9yes
21Claude Sonnet 4.6Anthropicmaxstandardv2026.06.248.341.5yes
22MiniMax M3MiniMaxenabledstandardv2026.06.244.622.3yes
23Kimi 2.6Moonshot AIenabledstandardv2026.06.244.622.1yes
24Qwen 3.7-PlusAlibabathinkingstandardv2026.06.242.821.5yes

Note: Default leaderboard view: full 108-task set, 500-step budget, foundation end-to-end GUI models. Results span several task releases (v2026.06.24, v2026.08.08, v2.1), which are not directly comparable; the site mixes them by default. Partial score = share of checkpoints passed. Organization mapped from model family as the site does.

Source: OSWorld 2.0 team (XLANG Lab, University of Hong Kong) · osworld-v2.xlang.ai · results as of 2026-11-14 · retrieved 2026-11-17

Markdown · More Agents, tool use & computer use