SShogo Bench

OSWorld

The original 369-task computer-use benchmark.

Agent Index weight
4%
in the Computer use pillar
Models tested
9
from the benchmark’s leaderboard and reports
Normalised for the index
0 → 0, 1 → 100
Official site

Results

#ModelScoreNormalisedHarnessCost/taskDate
1Claude Sonnet 4.6 Supported
Anthropic
72.1%claude-sonnet-4-6 (100 steps)——
2Claude Opus 4.5 (latest) Supported
Anthropic
66.3%Claude Opus 4.5—Nov 25, 2025
3Kimi K2.5 Supported
Moonshot AI
63.3%Kimi-K2.5—Feb 3, 2026
4Claude Sonnet 4.5 (latest) Supported
Anthropic
62.9%claude-sonnet-4-5-20250929 (100 steps)—Oct 30, 2025
5Claude Sonnet 4 Supported
Anthropic
43.9%claude-4-sonnet-20250514 (50 steps)—Jul 27, 2025
6Claude 3.7 Sonnet Supported
Anthropic
35.8%claude-3-7-sonnet-20250219 (50 steps)—Jul 27, 2025
7computer-use-preview-2025-03-11
Unknown
31.3%computer-use-preview (50 steps)—Jul 27, 2025
8o3 Supported
OpenAI
23.0%o3 (100 steps)—Jul 27, 2025
9Qwen2.5-72B Estimated
Alibaba (Qwen)
5.0%qwen2.5-vl-72b-instruct (100 steps)—Jul 27, 2025

For models without a OSWorld result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 8 models (residual SD 0.59 on the logit scale). Estimates are labelled and widen the model’s range.