The robots aren't coming for the lab benches yet, but OpenAI just built the test to see when they will.

The Summary

  • OpenAI released BioBasicsBench, a 100-task benchmark testing whether AI agents can handle routine wet-lab work like interpreting gel images, calculating dilutions, and designing experiments
  • Current frontier models (o1, o3-mini) score 40-60% on tasks that trained lab techs nail 95%+ of the time
  • The gap isn't knowledge, it's judgment: knowing when a Western blot failed versus when it's good enough to proceed

The Signal

Laboratory science has stayed stubbornly analog while software ate everything else. You still need humans to look at a gel, squint at the bands, and decide if the DNA extracted cleanly enough to sequence. You need someone to calculate serial dilutions without poisoning the cell culture. You need judgment calls that textbooks don't teach.

BioBasicsBench measures exactly that. The 100 tasks span the unglamorous daily work of molecular biology: interpreting experimental readouts, troubleshooting failed preps, designing basic protocols. OpenAI built it because everyone keeps asking when AI will automate lab work, but nobody had a real test for whether models could actually do it.

"The benchmark focuses on routine competence, not breakthrough discovery. Can an agent tell you why your PCR didn't work?"

The current results:

  • o1 scores 58% overall
  • o3-mini hits 52%
  • Claude and Gemini cluster around 40-45%
  • Human PhD students with 2+ years bench experience: 94%

The gap isn't factual knowledge. Models ace the theoretical questions about DNA replication or protein folding. They fall apart on applied judgment. Show them a gel photo and ask if the extraction worked well enough to proceed, they hedge or hallucinate. Ask them to troubleshoot a failed Western blot, they list 15 possible causes without ranking likelihood. Give them a cell culture contamination photo, they can't distinguish "start over" from "salvageable if you act now."

This matters because wet-lab bottlenecks are everywhere. Pharma companies spend months optimizing antibody production protocols that an experienced tech could tune in weeks. Academic labs burn grad student years on routine molecular cloning that requires zero creativity, just careful execution. Diagnostic labs process thousands of samples daily with judgment calls on every ambiguous readout.

The benchmark design reveals what's hard:

  • Image interpretation tasks (gel bands, microscopy, plate readers): 35% model accuracy
  • Protocol design from scratch: 42%
  • Troubleshooting failed experiments: 48%
  • Calculations and dilution planning: 71%

The models can calculate. They struggle with perception and context. A blurry band on a gel might mean degraded RNA or just uneven loading. An experienced human knows which follow-up question to ask. The model wants more data before committing, which in a lab setting means wasting another day and $200 in reagents.

The Implication

This benchmark is a roadmap, not a scorecard. The 40-point gap between models and humans marks exactly where the work needs to happen. Better vision models won't close it alone. Neither will bigger context windows. You need models that can commit to a decision with incomplete information and explain their reasoning well enough that a human can verify without redoing the whole experiment.

Watch for lab automation companies to use BioBasicsBench as a development target. The first model to hit 85% will unlock a different kind of agent: not a lab replacement, but a force multiplier that handles routine decisions and flags the genuinely ambiguous cases for human review. That's when bio research accelerates. Not because AI discovered anything, but because it freed up humans to focus on questions instead of quality control.

Sources

OpenAI Blog