Benchmarks · Last checked 25 Oct 2026
Computer-use benchmarks, explained
Benchmarks test agents on real tasks in real software. Scores rose fast in 2025–2026, but leaderboards often disagree. Here's what each test measures and how to read the numbers.
The headline numbers
42% → ~86%
Best OSWorld-Verified score, early 2025 to late 2026.
~72%
Human baseline on OSWorld-Verified (72.36%).
~44%
Best full-completion score on the harder OSWorld 2.0.
Latest reported scores
| Agent or model | Benchmark | Score | Reported by, date |
|---|---|---|---|
| Intelligence-Indeed Agent | OSWorld-Verified | 90.19% | Presenc, 1 Oct 2026 |
| Claude Fable 5 | OSWorld-Verified | 86.0% | Developers Digest, 17 Oct 2026 |
| Qwen3.8 Max | OSWorld-Verified | 86.1% | BenchLM, 22 Aug 2026 |
| Claude Opus 5 (max effort) | OSWorld 2.0 | ~44% full completion | Yutori, 2026 |
| Agent S3 (open source) | OSWorld | 69.9% | Verdent, Oct 2025 |
| OpenCUA-72B (open model) | OSWorld-Verified | 45.0% | NeurIPS 2025 |
These come from different sources and dates and may use different setups, so don't compare rows directly. Agents (a model plus scaffolding) often score higher than the same model alone.
What each benchmark measures
- OSWorld / OSWorld-Verified: about 369 tasks on real desktop operating systems and apps (office suites, browsers, file managers), checked automatically. "Verified" is a cleaned-up version with fixed tasks. The standard reference for desktop agents.
- OSWorld 2.0: a harder successor with longer tasks; the gap to human performance is much wider.
- WebArena: tasks on self-hosted copies of realistic websites (shopping, forums, code hosting), so results are reproducible.
- Online-Mind2Web: tasks on live public websites, closer to real use but noisier, because sites change.
Why leaderboards disagree
- Agent vs model: some entries are a raw model, others a full agent system with extra tools and retries.
- Different settings: step limits, screenshot resolution and "effort" settings change scores.
- Snapshot dates: aggregators update at different times.
- Who runs it: some leaderboards are run by vendors who also sell agents or infrastructure.
For buying decisions, a benchmark tells you which agents are worth a trial. Your own workflow, run 20–50 times, tells you whether one is good enough.
Live leaderboards
- OSWorld official site
- Yutori leaderboards (OSWorld-Verified, OSWorld 2.0, Online-Mind2Web and more)
- Steel OSWorld leaderboard
- BenchLM