OSWorld 2.1 (all 108 tasks) (OSWorld 2.1)
We show this table for reference; we do not rank on it.
OSWorld 2.1 is the current release of OSWorld 2.0: the same 108 long-horizon computer-use workflows with the maintainers' later fixes to task files, assets, mocked websites, and evaluators. This page stores partial scores on the full task set.
Previous release
OSWorld 2.0 results on the June and August 2026 releases remain on their own page. Version 2.1 fixes evaluators and task assets, so movement between the two tables is not a like-for-like model comparison. The 82-task offline subset has a separate page.
View OSWorld 2.0Benchmark score on OSWorld 2.1 — October 27, 2026
We compile the OSWorld 2.1 rows from provider self-reports. Claude Opus 5.5 leads the table at 81.8%, followed by Claude Fable 5.1 (80.7%) and Claude Sonnet 5.5 (80.1%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
Claude Fable 5.1
Anthropic
Claude Sonnet 5.5
Anthropic
9 modelsAgenticCurrentDisplay onlyUpdated October 27, 2026
Benchmark score table (9 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 81.8% | Not reported | Closed |
| 2 | Claude Fable 5.1Anthropic | 80.7% | Not reported | Closed |
| 3 | Claude Sonnet 5.5Anthropic | 80.1% | Not reported | Closed |
| 4 | Kimi K3.1 ProMoonshot AI | 79.6% | Not reported | Pending |
| 5 | Muse Wonder FullMeta | 78.4% | Not reported | Closed |
| 6 | GPT-6 AstraOpenAI | 75.8% | Not reported | Closed |
| 7 | Claude Opus 5Anthropic | 74.0% | Not reported | Closed |
| 8 | Muse WonderMeta | 72.6% | Not reported | Closed |
| 9 | Claude Sonnet 5Anthropic | 57.0% | Not reported | Closed |
Among the reported OSWorld 2.1 rows, Claude Opus 5.5 is first at 81.8%. The third row is 1.7 points behind. The broader top-10 range is 24.8 points, so the table still separates the published systems.
9 models have been evaluated on OSWorld 2.1. The benchmark falls in the Agentic category. OSWorld 2.1 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About OSWorld 2.1
Year
2026
Tasks
108 long-horizon computer-use workflows
Format
Interactive computer-use evaluation; partial score
Difficulty
Long-horizon professional workflows
The osworld-v2.1 release pins task files, gated assets, mocked websites, and the Ubuntu VM image for reproducible runs and adds an offline task subset. Labs report a partial score (mean per-task checkpoint credit) and a strict pass rate; we store the partial score, the metric every source here publishes, and keep strict pass rates in the source notes. Runs use each lab's own harness, so treat gaps of a few points as unresolved. Results on the earlier OSWorld 2.0 releases stay on the OSWorld 2.0 page, and the 82-task offline subset has its own page.
Freshness and provenance
Version
OSWorld 2.1 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
How is OSWorld 2.1 different from OSWorld 2.0?
OSWorld 2.1 is the maintainers' current release of OSWorld 2.0. It keeps the 108 workflows but fixes task evaluators, assets, and mocked websites, so its scores are not like-for-like with results on the June or August 2026 OSWorld 2.0 releases. Those results stay on the OSWorld 2.0 page.
Why is the offline subset on a separate page?
The offline subset runs 82 of the 108 tasks with no internet access in the agent's virtual machine. Anthropic reports Claude Sonnet 5.5 at 80.1% on all 108 tasks and 83.9% on the offline subset, so mixing the two would distort the comparison.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.