OSWorld-Verified 2.0
Computer-use tasks in a real desktop environment.
As of 24 Oct 2026, Claude Opus 5 leads OSWorld-Verified 2.0 on BenchLeader with 31.4%, ahead of Claude Opus 5 at 30.2%, across 14 model configurations with a published result.
- Published by
- OSWorlddata via Epoch AI Benchmarking Hub
- Category
- Agents & tools
- Index weight
- 0.5
- Models
- 14
- Data as of
- 24 Oct 2026
- Newest result
- 24 Jul 2026
The publisher has posted no new results on this board since 24 Jul 2026. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
OSWorld-Verified gives an agent a real Ubuntu desktop with office suites, browsers and file managers, and asks it to complete tasks a human would do with mouse and keyboard, judged by checking the resulting files and application state.
How it is scored
Percent of tasks completed, on the verified task set with corrected checkers, at a fixed step budget.
What to keep in mind
Scores depend on screenshot resolution, step limits and the computer-use harness, so compare runs with the same settings. Human success on the same tasks is around 70%, which puts model scores in context.
- 1Claude Opus 5 (max)31.4%
- 2Claude Opus 5 (xhigh)30.2%
- 3Claude Opus 5 (high)29.0%
- 4GPT-5.6 Sol (max)27.3%
- 5Claude Opus 5 (medium)25.3%
- 6Claude Opus 5 (low)22.3%
- 7Claude Opus 4.8 (max)20.6%
- 8Claude Opus 4.7 (max)18.2%
- 9GPT-5.5 (xhigh)13.0%
- 10Claude Sonnet 4.6 (medium)9.3%
- 11Claude Sonnet 4.6 (max)8.3%
- 12MiniMax M34.6%
- 13Kimi K2.64.6%
- 14Qwen3.7 Plus2.8%
| # | ||||
|---|---|---|---|---|
| 1 |  Claude Opus 5max | 31.4% | 68.8 | 2026-07-24 |
| 2 | Claude Opus 5xhigh | 30.2% | 67.4 | 2026-07-24 |
| 3 | Claude Opus 5high | 29.0% | 67.9 | 2026-07-24 |
| 4 |  GPT-5.6 Solmax | 27.3% | 68.0 | 2026-07-09 |
| 5 | Claude Opus 5medium | 25.3% | 64.6 | 2026-07-24 |
| 6 | Claude Opus 5low | 22.3% | 61.3 | 2026-07-24 |
| 7 | Claude Opus 4.8max | 20.6% | 63.3 | 2026-05-28 |
| 8 | Claude Opus 4.7max | 18.2% | 62.4 | 2026-04-16 |
| 9 | GPT-5.5xhigh | 13.0% | 66.1 | 2026-04-23 |
| 10 | Claude Sonnet 4.6medium | 9.3% | – | 2026-02-17 |
| 11 | Claude Sonnet 4.6max | 8.3% | 56.0 | 2026-02-17 |
| 12 |  MiniMax M3open ↗ | 4.6% | 56.4 | 2026-06-01 |
| 13 |  Kimi K2.6open ↗ | 4.6% | 55.4 | 2026-04-20 |
| 14 |  Qwen3.7 Plus | 2.8% | 58.5 | 2026-06-02 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “OSWorld-Verified 2.0 leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/osworld_2.
OSWorld-Verified 2.0: questions
- What does OSWorld-Verified 2.0 measure?
- OSWorld-Verified gives an agent a real Ubuntu desktop with office suites, browsers and file managers, and asks it to complete tasks a human would do with mouse and keyboard, judged by checking the resulting files and application state. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads OSWorld-Verified 2.0?
- Claude Opus 5 leads OSWorld-Verified 2.0 with 31.4% as of 24 Oct 2026, ahead of Claude Opus 5 at 30.2%.
- How many models have OSWorld-Verified 2.0 results?
- 14 model configurations have a OSWorld-Verified 2.0 result on BenchLeader, all taken from OSWorld via Epoch AI Benchmarking Hub.
- Who runs OSWorld-Verified 2.0 and how often is it updated?
- OSWorld-Verified 2.0 is published by OSWorld. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does OSWorld-Verified 2.0 count toward the BenchLeader Index?
- Yes. OSWorld-Verified 2.0 contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.