AIEvals

Benchmarks

OSWorld 2.1 (offline strict pass)

Completing every checkpoint in a long-horizon desktop workflow on the official 82-task offline OSWorld 2.1 subset.

Computer useReported by AnthropicVendor-reported onlyReported by Anthropic

How it is run

Section 8.9.3 reports strict pass@1 averaged over five attempts per task, at 1080p with a 500-step limit and no virtual-machine internet access.

What it cannot tell you

Strict pass counts only fully completed attempts. It must not share ranks with partial checkpoint credit or earlier combined online/offline task sets.

Why there is no independent table

No independent suite in this dataset runs OSWorld 2.1 (offline strict pass). The numbers below are the ones model publishers report about their own models, which is a different kind of evidence: each publisher chooses its own harness and effort level, so a row is readable on its own and not against the other rows.

Read this with

Other versions of the same evaluation. Versions are not comparable: the task set changes between them.

  • OSWorld 2.0Same family. Whether an agent can carry out a long-horizon computer-use workflow on a real desktop: opening applications, moving files, filling in software.
  • OSWorld 2.1Same family. Whether an agent can carry out a long-horizon computer-use workflow on a real desktop: opening applications, moving files, filling in software, scored as partial reward on the offline set.
  • CUA-benchSame category. Playing six held-out commercial games through keyboard and mouse interaction.
  • Agents' Last ExamSame category. Complex professional agent workflows carried out end to end, rather than a single question answered correctly.

As reported by each publisher

Kept separate per publisher, because the settings are not shared.

Anthropic

Claude Haiku 5.5: system card capability evaluations · 7 October 2026

RankModelScore
4Claude Haiku 5.537.1%
6Claude Haiku 4.51.7%
2Claude Sonnet 5.548.8%
1Claude Opus 5.553.2%
5GPT-6 Luna17.1%
3GPT-6.1 Sol39.8%

Ranks compare reported scores in the same published row; ties share a rank.

OSWorld 2.1 (offline strict pass). §8.9.3 / Figure 8.9.3.A: every checkpoint must pass. Same 82-task offline setup; distinct from partial credit.

Section 8, pages 111–139 of the Haiku 5.5 system card, read 8 October 2026. Values come from exact prose, tables and labelled figures. Unless specified, Haiku uses adaptive thinking at max effort and five trials with evaluation-specific context up to 1M. Each row retains its subset, metric, effort and source notes; no score is assigned to an unspecified SWE-Bench Pro version. Terminal runs use safeguards without fallback and no internet egress. ProgramBench uses filtered tasks and hidden-test pass rate without a time limit. OSWorld partial and strict results use the official 82-task offline subset. CAD uses 0–1 voxel IoU. Healthcare raw and length-adjusted scores stay separate. Life-science results disable Haiku biology safeguards. Independently run figures quoted by Anthropic remain publisher reports here. Missing comparison values are not inferred from rounded charts.

www.anthropic.com/claude-haiku-5-5-system-card

Sources

Every page used for this benchmark, read on 17 October 2026.