LabOSBench - Experimental Execution: leaderboard
Experimental-execution operations in LabOSBench (Zou et al., 2026), such as starting scans and captures, running milling and deposition patterns, and waiting for an experiment to complete, across its web-based scientific-instrument simulators. Each subtask starts from its canonical pre-state with a 50-step budget and runs twice; the score is the percent of successful runs averaged over the category's subtasks; higher is better.
Metric: Subtask Success Rate (%; mean over the category's operation-level subtasks, 2 runs each). Source: arxiv.org. Saturation forecast: Around 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed-1.6 | 82.1 | #447 |
| 2 | GPT-5.5 | 71.4 | #48 |
| 3 | Kimi K2.5 | 57.1 | #252 |
| 4 | Claude Sonnet 4.5 | 35.7 | #244 |
| 5 | Claude Opus 4.5 | 35.7 | #152 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=labosbench-experimental-execution · How It Works · Data refreshed daily, snapshot 2026-11-12.