Benchmark · Multimodal
OSWorld 2.0
% tasks completed (%) Higher is better Benchmark homepage
What it measures
OSWorld 2.0 tests multimodal computer-use agents on 108 long-horizon computer-use workflows spanning everyday and professional tasks. It is developed by the XLANG Lab at the University of Hong Kong and co-authors. The primary metric is binary task completion; a partial score based on task-specific checkpoints is reported alongside it. There are three task releases (v2026.06.24, v2026.08.08 and v2.1, the recommended one) and an 82-task offline set that runs without internet access.
Caveats:
Numbers are only comparable when release, task set (full or offline), metric (binary or partial), step budget (150, 300 or 500) and agent approach (standard or batch tool calls) all match. Providers usually quote partial scores, which are much higher than binary completion. Claude Opus 5 at max effort on v2.1 scores 44.33% binary and 77.67% partial. The verified leaderboard is run by the maintainers and covers few current models, so most numbers are self-reported. The original OSWorld and OSWorld-Verified (369 tasks) are a different benchmark and are not recorded here.
Results
| Model | Score | Setting | Reported by | Measured | Verified | Source |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 Anthropic | 87.2 % | v2.1, offline subset (82 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| Claude Sonnet 5.5 Anthropic | 83.9 % | v2.1, offline subset (82 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| Claude Opus 5.5 Anthropic | 81.8 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Opus 5.5 System Card (Anthropic, 2026-09-22) |
| Claude Fable 5.1 Anthropic | 80.7 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Opus 5.5 System Card (Anthropic, 2026-09-22) |
| Claude Sonnet 5.5 Anthropic | 80.1 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Sonnet 5.5 System Card (Anthropic, 2026-09-28) |
| Muse Wonder Full Meta | 78.4 % | v2.1, full 108-task set, partial (mean per-task score, primary metric); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| Claude Fable 5.1 Anthropic | 77.9 % | August 2026 task release, full set (108 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Fable 5.1 & Claude Mythos 5.1 System Card (Anthropic, 2026-09-01) |
| GPT-6 Astra OpenAI | 72.6 % | v2026.08.08, offline set, partial score; max effort | Self-reported | — | 2026-10-25 | GPT-6 Astra: A new generation of intelligence | OpenAI |
| Muse Wonder Meta | 72.6 % | v2.1, full 108-task set, partial (mean per-task score, primary metric); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| Claude Haiku 5.5 Anthropic | 72.4 % | v2.1, offline subset (82 tasks), partial; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| GPT-6 Astra OpenAI | 71.7 % | v2026.08.08, offline set, partial score; xhigh effort | Self-reported | — | 2026-10-25 | GPT-6 Astra: A new generation of intelligence | OpenAI |
| GPT-6 Astra OpenAI | 71.6 % | v2026.08.08, offline set, partial score; high effort | Self-reported | — | 2026-10-25 | GPT-6 Astra: A new generation of intelligence | OpenAI |
| GPT-6.1 Sol OpenAI | 71.4 % | v2026.08.08, offline set, partial score; max effort | Self-reported | — | 2026-10-25 | Introducing GPT-6.1 Sol | OpenAI |
| GPT-6.1 Sol OpenAI | 69.6 % | v2026.08.08, offline set, partial score; high effort | Self-reported | — | 2026-10-25 | Introducing GPT-6.1 Sol | OpenAI |
| Muse Spark 1.3 Meta | 69.56 % | v2.1, full set, partial; batch tool approach, 500-step budget; max effort | Third party | 2026-10-15 | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| GPT-6.1 Sol OpenAI | 69.4 % | v2026.08.08, offline set, partial score; xhigh effort | Self-reported | — | 2026-10-25 | Introducing GPT-6.1 Sol | OpenAI |
| Gemini 4 Argon Google | 69.2 % | v2026.08.08, offline set, partial score; high effort; batch tool; max 500 steps, best of 3 runs | Self-reported | — | 2026-10-25 | Gemini 4 Argon model card - Google DeepMind |
| Muse Spark 1.3 Meta | 66.9 % | version 08.08, full 108-task set, partial (mean per-task score, primary metric); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 model page |
| GPT-6.1 Sol OpenAI | 66.8 % | v2026.08.08, offline set, partial score; medium effort | Self-reported | — | 2026-10-25 | Introducing GPT-6.1 Sol | OpenAI |
| GPT-6 Sol OpenAI | 64.4 % | v2026.08.08, offline set, partial score; max effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| GPT-6 Astra OpenAI | 61.4 % | v2026.08.08, offline set, partial score; low effort | Self-reported | — | 2026-10-25 | GPT-6 Astra: A new generation of intelligence | OpenAI |
| GPT-6 Sol OpenAI | 60.5 % | v2026.08.08, offline set, partial score; xhigh effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Gemini 3.8 Flash Google | 59 % | release before the 2026-08-08 patch, set not specified, partial score; batch tool; effort not stated; max 500 steps, best of 3 runs | Self-reported | — | 2026-10-25 | Gemini 3.8 Flash model card - Google DeepMind |
| GPT-6.1 Sol OpenAI | 59 % | v2026.08.08, offline set, partial score; low effort | Self-reported | — | 2026-10-25 | Introducing GPT-6.1 Sol | OpenAI |
| Muse Spark 1.3 Meta | 59 % | version 08.08, full 108-task set, partial; xhigh effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 Evaluation Methodology (PDF) |
| GPT-6 Sol OpenAI | 58.3 % | v2026.08.08, offline set, partial score; high effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| GPT-6 Sol OpenAI | 54 % | v2026.08.08, offline set, partial score; medium effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Claude Opus 5.5 Anthropic | 53.2 % | v2.1, offline subset (82 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| GPT-6 Luna OpenAI | 52.7 % | v2026.08.08, offline set, partial score; max effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Claude Sonnet 5.5 Anthropic | 48.8 % | v2.1, offline subset (82 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| Claude Opus 5.5 Anthropic | 48.7 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Opus 5.5 System Card (Anthropic, 2026-09-22) |
| Qwen3.8-27B Alibaba Cloud (Qwen) | 48 % | release and task set not stated, partial; effort not stated | Self-reported | — | 2026-10-25 | Qwen/Qwen3.8-Flash-Next - Hugging Face model card |
| Muse Spark 1.2 Meta | 47.6 % | version 06.24, full 108-task set, partial; xhigh effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 model page |
| Muse Spark 1.1 Meta | 47.3 % | release not stated, full 108-task set, partial; xhigh effort; GUI computer-control tool only ('w/o exec'), max 500 steps | Self-reported | — | 2026-10-25 | Muse Spark 1.1 Evaluation Report (Meta, 9 Jul 2026) |
| GPT-6 Luna OpenAI | 46.7 % | v2026.08.08, offline set, partial score; xhigh effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Qwen3.8-Max Alibaba Cloud (Qwen) | 46.7 % | release and task set not stated, partial; effort not stated | Self-reported | — | 2026-10-25 | Qwen3.8-Max: A New Bar for Coding and Cowork (Qwen blog, 2026-08-03) |
| Muse Spark 1.2 Meta | 44.08 % | v2.1, full set, partial; batch tool approach, 500-step budget; xhigh effort | Third party | 2026-10-15 | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| GPT-6 Sol OpenAI | 43.9 % | v2026.08.08, offline set, partial score; low effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Claude Sonnet 5.5 Anthropic | 43.5 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Sonnet 5.5 System Card (Anthropic, 2026-09-28) |
| Claude Fable 5.1 Anthropic | 42.8 % | v2.1 (Sep 10, 2026 task files), full set (108 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Opus 5.5 System Card (Anthropic, 2026-09-22) |
| Claude Fable 5.1 Anthropic | 41.7 % | August 2026 task release, full set (108 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Fable 5.1 & Claude Mythos 5.1 System Card (Anthropic, 2026-09-01) |
| GPT-6 Luna OpenAI | 41.4 % | v2026.08.08, offline set, partial score; high effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Muse Wonder Full Meta | 39.4 % | v2.1, full 108-task set, binary (strict completion); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| Muse Spark 1.3 Meta | 39.25 % | v2.1, full set, binary; batch tool approach, 500-step budget; max effort | Third party | 2026-10-15 | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| Claude Haiku 5.5 Anthropic | 37.1 % | v2.1, offline subset (82 tasks), strict; max effort; 500-step budget | Self-reported | — | 2026-10-12 | Claude Haiku 5.5 System Card (Anthropic, 2026-10-07) |
| Muse Wonder Meta | 34.3 % | v2.1, full 108-task set, binary (strict completion); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| Muse Spark 1.3 Meta | 32 % | version 08.08, full 108-task set, binary (strict completion); max effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 model page |
| GPT-6 Luna OpenAI | 31.5 % | v2026.08.08, offline set, partial score; medium effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Muse Spark 1.3 Meta | 26.7 % | version 08.08, full 108-task set, binary; xhigh effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 Evaluation Methodology (PDF) |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 21.5 % | release and task set not stated, partial; effort not stated | Self-reported | — | 2026-10-25 | Qwen3.8-Max: A New Bar for Coding and Cowork (Qwen blog, 2026-08-03) |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 21.5 % | v2026.06.24, full set, partial; standard approach, 500-step budget; thinking | Third party | — | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| Qwen3.8-27B Alibaba Cloud (Qwen) | 19.4 % | release and task set not stated, binary; effort not stated | Self-reported | — | 2026-10-25 | Qwen/Qwen3.8-Flash-Next - Hugging Face model card |
| Qwen3.8-Max Alibaba Cloud (Qwen) | 19.4 % | release and task set not stated, binary; effort not stated | Self-reported | — | 2026-10-25 | Qwen3.8-Max: A New Bar for Coding and Cowork (Qwen blog, 2026-08-03) |
| Muse Spark 1.2 Meta | 17.9 % | version 06.24, full 108-task set, binary; xhigh effort; GUI computer-control tool in Meta's internal evaluation framework | Self-reported | — | 2026-10-25 | Muse Spark 1.3 model page |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 16.6 % | v2026.06.24, full set, partial; standard approach, 300-step budget; thinking | Third party | — | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| Muse Spark 1.2 Meta | 15.84 % | v2.1, full set, binary; batch tool approach, 500-step budget; xhigh effort | Third party | 2026-10-15 | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| Muse Spark 1.1 Meta | 14.2 % | release not stated, full 108-task set, binary; xhigh effort; GUI computer-control tool only ('w/o exec'), max 500 steps | Self-reported | — | 2026-10-25 | Muse Spark 1.1 Evaluation Report (Meta, 9 Jul 2026) |
| GPT-6 Luna OpenAI | 8.3 % | v2026.08.08, offline set, partial score; low effort | Self-reported | — | 2026-10-25 | Introducing GPT-6 Sol and Luna | OpenAI |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 2.8 % | release and task set not stated, binary; effort not stated | Self-reported | — | 2026-10-25 | Qwen3.8-Max: A New Bar for Coding and Cowork (Qwen blog, 2026-08-03) |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 2.8 % | v2026.06.24, full set, binary; standard approach, 500-step budget; thinking | Third party | — | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
| Qwen3.7-Plus Alibaba Cloud (Qwen) | 1.9 % | v2026.06.24, full set, binary; standard approach, 300-step budget; thinking | Third party | — | 2026-10-25 | OSWorld 2.0 (project page and leaderboard) |
Sorted by score, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source. Scores are only directly comparable between rows with the same setting. Self-reported results come from the provider and have not been reproduced independently.
Sources
- OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks (arXiv 2606.29537)
- OSWorld 2.0 (project page and leaderboard)
- Claude Haiku 5.5 System Card (Anthropic, 2026-10-07)
- Claude Opus 5.5 System Card (Anthropic, 2026-09-22)
- Claude Sonnet 5.5 System Card (Anthropic, 2026-09-28)
- Muse Wonder Model Card (PDF)
- Claude Fable 5.1 & Claude Mythos 5.1 System Card (Anthropic, 2026-09-01)
- GPT-6 Astra: A new generation of intelligence | OpenAI
- Introducing GPT-6.1 Sol | OpenAI
- Gemini 4 Argon model card - Google DeepMind
- Muse Spark 1.3 model page
- Introducing GPT-6 Sol and Luna | OpenAI
- Gemini 3.8 Flash model card - Google DeepMind
- Muse Spark 1.3 Evaluation Methodology (PDF)
- Qwen/Qwen3.8-Flash-Next - Hugging Face model card
- Muse Spark 1.1 Evaluation Report (Meta, 9 Jul 2026)
- Qwen3.8-Max: A New Bar for Coding and Cowork (Qwen blog, 2026-08-03)