Computer use · %
OSWorld (Verified)
369 real computer tasks in a full Ubuntu desktop (browsers, office apps, file manager, terminal) checked by execution-based scripts.
What it measures
Task success rate of computer-use agents. Data compiled by Epoch AI from the OSWorld site.
Limitations
Step budgets (15/50/100) and agent frameworks differ between rows; compare like with like.
Contamination risk: Medium
Task definitions are public.
Leaderboard · Subset / setting: compiled
1 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.5 | Anthropic | 66.3% | 1 | OSWorld; agent: Claude Opus 4.5; effort: no thinking | Vendor-reported via Epoch AI | Nov 25, 2025 |
Leaderboard · Subset / setting: leaderboard
8 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | Anthropic | 72.1% | 1 | OSWorld; agent: claude-sonnet-4-6 (100 steps); effort: no thinking | Official leaderboard via Epoch AI | Feb 17, 2026 |
| 2 | Kimi K2.5OW | Moonshot AI | 63.3% | 2 | OSWorld; agent: Kimi-K2.5 | Official leaderboard via Epoch AI | Feb 3, 2026 |
| 3 | Claude Sonnet 4.5 | Anthropic | 62.9% | 3 | OSWorld; agent: claude-sonnet-4-5-20250929 (100 steps); effort: no thinking | Official leaderboard via Epoch AI | Oct 30, 2025 |
| 4 | Claude Sonnet 4 | Anthropic | 43.9% | 5 | OSWorld; agent: claude-4-sonnet-20250514 (50 steps) | Official leaderboard via Epoch AI | Jul 27, 2025 |
| 5 | Claude 3.7 Sonnet | Anthropic | 35.8% | 8 | OSWorld; agent: claude-3-7-sonnet-20250219 (50 steps) | Official leaderboard via Epoch AI | Jul 27, 2025 |
| 6 | computer-use-preview-2025-03-11 | — | 31.3% | 10 | OSWorld; agent: computer-use-preview (50 steps) | Official leaderboard via Epoch AI | Jul 27, 2025 |
| 7 | o3 | OpenAI | 23.0% | 15 | OSWorld; agent: o3 (100 steps); effort: medium | Official leaderboard via Epoch AI | Jul 27, 2025 |
| 8 | Qwen2.5-72BOW | Alibaba (Qwen) | 5.0% | 18 | OSWorld; agent: qwen2.5-vl-72b-instruct (100 steps) | Official leaderboard via Epoch AI | Jul 27, 2025 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2025-07-2743.9%Claude Sonnet 4
- 2025-10-3062.9%Claude Sonnet 4.5
- 2026-02-0363.3%Kimi K2.5
- 2026-02-1772.1%Claude Sonnet 4.6
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| Epoch AI Benchmarking Hub Epoch AI, 'Capabilities & benchmarking', https://epoch.ai/benchmarks (CC BY 4.0) | CC BY 4.0 | Open licence – full data with attribution | Oct 22, 2026 |
FAQ
What does OSWorld (Verified) measure?
369 real computer tasks in a full Ubuntu desktop (browsers, office apps, file manager, terminal) checked by execution-based scripts. Task success rate of computer-use agents. Data compiled by Epoch AI from the OSWorld site.
How reliable is OSWorld (Verified)?
Step budgets (15/50/100) and agent frameworks differ between rows; compare like with like. Contamination risk: Medium – Task definitions are public.