LabOSBench: leaderboard
Computer-use benchmark for scientific-instrument software (LabOSBench, Zou et al., 2026): GUI agents operate eight web-based simulators (scanning and transmission electron microscopes, light microscope, X-ray diffraction, focused ion beam, atom probe, energy-dispersive spectroscopy and scanning probe) through 96 operation-level subtasks, each started from its canonical pre-state with a 50-step budget and run twice. The score is the subtask success rate averaged per instrument and then over the eight instruments, in percent; higher is better.
Metric: Subtask Success Rate (%; subtask-averaged per instrument, macro-averaged over 8 instruments, 2 runs per subtask). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed-1.6 | 76.3 | #447 |
| 2 | GPT-5.5 | 72.6 | #48 |
| 3 | Claude Sonnet 4.5 | 62.4 | #244 |
| 4 | Claude Opus 4.5 | 62 | #152 |
| 5 | Kimi K2.5 | 59.7 | #252 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=labosbench · How It Works · Data refreshed daily, snapshot 2026-11-12.