Best LLM for computer use
- Claude Opus 5 leads
- 31.4%
- models ranked
- 9
- lead over #2
- +4.1 pts
OSWorld 2.0 · read October 21
# OSWorld 2.0, read October 21
leaders = {
"Claude Opus 5": 31.4%, "GPT-5.6 Sol": 27.3%, "Claude Opus 4.8": 20.6%,
}
As of October 21, 2026, Anthropic Claude Opus 5 leads 9 models on OSWorld 2.0 with 31.4%, ahead of OpenAI GPT-5.6 Sol (27.3%) and Anthropic Claude Opus 4.8 (20.6%).
This list narrows the best LLM for agents ranking.
| Rank | Model | OSWorld 2.0 | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Opus 5 Anthropic | 31.4% | 81.8 | $25.00 | |
| 2 | GPT-5.6 Sol OpenAI | 27.3% | 76.7 | $20.00 | |
| 3 | Claude Opus 4.8 Anthropic | 20.6% | 66.0 | $25.00 | |
| 4 | Claude Opus 4.7 Anthropic | 18.2% | 51.1 | $25.00 | |
| 5 | GPT-5.5 OpenAI | 13.0% | 67.8 | $30.00 | |
| 6 | Claude Sonnet 4.6 Anthropic | 9.3% | 57.8 | $15.00 | |
| 7 | MiniMax M3 MiniMax | 4.6% | 57.8 | $1.20 | |
| 8 | Kimi K2.6 Moonshot AI | 4.6% | 44.2 | $3.41 | |
| 9 | Qwen3.7 Plus Qwen | 2.8% | Not benchmarked | $1.28 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Models by their OSWorld 2.0 result, as the board publishes it, best first. Tasks on a real desktop computer, done through the screen, mouse and keyboard. The number is the share of tasks completed.
- OpenRouter, fetched
- Epoch AI benchmarking hub, fetched