OSWorld 2.1 offline subset (82 tasks) (OSWorld 2.1 Offline)
We show this table for reference; we do not rank on it.
The official offline subset of OSWorld 2.1: 82 of the 108 long-horizon computer-use workflows, run with no internet access in the agent's virtual machine.
Full task set
This page covers the 82-task offline subset of OSWorld 2.1. Results on all 108 tasks are on the main OSWorld 2.1 page and are not comparable with these scores.
View OSWorld 2.1 (108 tasks)Benchmark score on OSWorld 2.1 Offline — October 29, 2026
We compile the OSWorld 2.1 Offline rows from provider self-reports. Claude Opus 5.5 leads the table at 87.2%, followed by Claude Sonnet 5.5 (83.9%) and Gemini 4 Carbon (81.5%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
Claude Sonnet 5.5
Anthropic
Gemini 4 Carbon
8 modelsAgenticCurrentDisplay onlyUpdated October 29, 2026
Benchmark score table (8 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 87.2% | Not reported | Closed |
| 2 | Claude Sonnet 5.5Anthropic | 83.9% | Not reported | Closed |
| 3 | Gemini 4 CarbonGoogle | 81.5% | Not reported | Closed |
| 4 | Gemini 4 ArgonGoogle | 74.7% | Not reported | Closed |
| 5 | GPT-6.1 TerraOpenAI | 74.2% | Not reported | Closed |
| 6 | Claude Haiku 5.5Anthropic | 72.4% | Not reported | Closed |
| 7 | GPT-6 LunaOpenAI | 48.9% | Not reported | Closed |
| 8 | Claude Haiku 4.5Anthropic | 15.7% | Not reported | Closed |
Among the reported OSWorld 2.1 Offline rows, Claude Opus 5.5 is first at 87.2%. The third row is 5.7 points behind. The broader top-10 range is 71.5 points, so the table still separates the published systems.
8 models have been evaluated on OSWorld 2.1 Offline. The benchmark falls in the Agentic category. OSWorld 2.1 Offline is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About OSWorld 2.1 Offline
Year
2026
Tasks
82 offline computer-use workflows
Format
Interactive computer-use evaluation; partial score
Difficulty
Long-horizon professional workflows
The osworld-v2.1 release added offline task subsets. Anthropic and OpenAI now report this 82-task offline list. We store the partial score (mean per-task checkpoint credit) and keep strict pass rates in the source notes. Scores are not comparable with the full 108-task OSWorld 2.1 page or with OSWorld 2.0.
Freshness and provenance
Version
OSWorld 2.1 Offline 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
What does OSWorld 2.1 Offline measure?
The official offline subset of OSWorld 2.1: 82 of the 108 long-horizon computer-use workflows, run with no internet access in the agent's virtual machine.
Which model scores highest on OSWorld 2.1 Offline?
Claude Opus 5.5 by Anthropic currently leads with a score of 87.2% on OSWorld 2.1 Offline.
How many models are evaluated on OSWorld 2.1 Offline?
8 AI models have published results on OSWorld 2.1 Offline in the BenchLM catalog.
Compare top models on OSWorld 2.1 Offline
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.