Gamow Labs has released LabRobotBench v0.1, a benchmark evaluating how well AI agents can operate wet-lab robots and recover from equipment failures. The suite tests frontier models on physical reasoning and complex biological workflows across three scenarios: restarting interrupted cell assays, juggling competing deadlines for cold storage and experiments, and completing time-sensitive dose-response tests.
Results show GPT-4.1 leading with a 46.7% mean recovery score, followed by Opus at 35.5%, Claude Code at 23.0%, and Gemini at 20.9%. In the first case, agents had to finish preparing three viable cell trays and discard a spoiled fourth; Opus returned an unfinished tray without adding the critical reagent. The second case forced agents to prioritize between finishing experiments within 200 seconds and returning cold supplies within 240 seconds — GPT-4.1 saved the supplies but missed an experiment, while Opus did the opposite. The third case required finishing a dose-response assay before a measurement deadline.
The benchmark aims to enable overnight, unattended experimentation by giving agents the ability to diagnose errors and resume work safely, potentially multiplying scientific throughput.
Source: Hacker News · Summarized by HeadlinesBriefing