Computer use · %
OSWorld 2.0
Long-horizon successor to OSWorld (2026) with multi-application workflows.
Active Contamination risk: Low 14 results · 9 models Updated Oct 23, 2026 Newest result: 2026-07-24 Official leaderboard ↗
What it measures
Binary task success (partial credit reported separately).
Limitations
Few systems evaluated so far; expensive to run.
Contamination risk: Low
New tasks.
Leaderboard
9 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | Anthropic | 31.4% | 1 | Binary accuracy; 500 steps; reasoning: max; effort: max | Official leaderboard via Epoch AI | Jul 24, 2026 |
| 2 | GPT-5.6 Sol | OpenAI | 27.3% | 4 | Binary accuracy; 500 steps; reasoning: max; effort: max | Official leaderboard via Epoch AI | Jul 9, 2026 |
| 3 | Claude Opus 4.8 | Anthropic | 20.6% | 7 | Binary accuracy; 500 steps; reasoning: max | Official leaderboard via Epoch AI | May 28, 2026 |
| 4 | Claude Opus 4.7 | Anthropic | 18.2% | 8 | Binary accuracy; 500 steps; reasoning: max; effort: max | Official leaderboard via Epoch AI | Apr 16, 2026 |
| 5 | GPT-5.5 | OpenAI | 13.0% | 9 | Binary accuracy; 500 steps; reasoning: xhigh; effort: xhigh | Official leaderboard via Epoch AI | Apr 23, 2026 |
| 6 | Claude Sonnet 4.6 | Anthropic | 9.3% | 10 | Binary accuracy; 500 steps; reasoning: medium; effort: medium | Official leaderboard via Epoch AI | Feb 17, 2026 |
| 7 | MiniMax-M3OW | MiniMax | 4.6% | 12 | Binary accuracy; 500 steps; reasoning: enabled | Official leaderboard via Epoch AI | Jun 1, 2026 |
| 8 | Kimi K2.6OW | Moonshot AI | 4.6% | 12 | Binary accuracy; 500 steps; reasoning: enabled | Official leaderboard via Epoch AI | Apr 20, 2026 |
| 9 | Qwen3.7-Plus | Alibaba (Qwen) | 2.8% | 14 | Binary accuracy; 500 steps; reasoning: thinking | Official leaderboard via Epoch AI | Jun 2, 2026 |
Best score over time
Each step is a new best result in our data, by evaluation date.
- 2026-02-179.3%Claude Sonnet 4.6
- 2026-04-1618.2%Claude Opus 4.7
- 2026-05-2820.6%Claude Opus 4.8
- 2026-07-0927.3%GPT-5.6 Sol
- 2026-07-2431.4%Claude Opus 5
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 23, 2026 |
FAQ
What does OSWorld 2.0 measure?
Long-horizon successor to OSWorld (2026) with multi-application workflows. Binary task success (partial credit reported separately).
How reliable is OSWorld 2.0?
Few systems evaluated so far; expensive to run. Contamination risk: Low – New tasks.