66.3%
▲ 61.3 pts since 2024
OSWorld accuracy of Claude Opus 4.5 (no thinking), 2025; human baseline 72.4%
20242025
What the chart shows
OSWorld puts AI agents in real Ubuntu, Windows and macOS environments and asks them to complete open-ended tasks across desktop and web apps, files and multi-app workflows. The best result rose from 5% (Qwen2.5-72B) in 2024 to 66.3% (Claude Opus 4.5, without extended thinking) in 2025.
Computer science students complete 72.4% of the tasks, so the gap to people has narrowed to about 6 percentage points. The AI Index calls this one of the fastest-closing gaps among its agent benchmarks.
Key numbers
| Measure | Value |
|---|---|
| 2024: Qwen2.5-72B | 5.0% |
| 2025: Claude Opus 4.5 (no thinking) | 66.3% |
| Human baseline (computer science students) | 72.35% |
| Gap to people, 2025 | 6.05 percentage points |
How to read it
- Data from Epoch AI’s benchmarking hub, as cited by the AI Index; each bar is the best result in that year.
- The benchmark has 369 tasks; students took a median of about two minutes per task, per the AI Index.
Source: Epoch AI, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.6.2: OSWorld: accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion