OpenAI Launches BioBasicsBench to Test AI Agents on Laboratory Tasks
Developed with Ligo Biosciences, the benchmark tests AI agents on 100 routine laboratory tasks involving experimental data, error detection and scientific reasoning.
Topics
News
- Cyber Breaches Hit 86% of Companies as AI Widens Security Skills Gap
- Anthropic Strikes $3.4 Billion Deal for Gas-Fired Power Generators
- OpenAI Launches BioBasicsBench to Test AI Agents on Laboratory Tasks
- Coforge Joins SpaceX Partner Network to Bring Grok AI to Enterprises
- Meta Warns US Midterm Backlash Could Slow Data Center Expansion
- Google to Invest $17 Billion in Wyoming Data Center Despite US Backlash
OpenAI has introduced BioBasicsBench, a benchmark developed with biotech company Ligo Biosciences to assess how reliably AI agents can interpret experimental data, identify errors and troubleshoot routine laboratory work.
Announced on November 6, the benchmark covers 100 tasks across five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.
The tasks use raw experimental evidence, including gel images, microscopy, sequencing files, instrument readouts and laboratory videos. They test whether AI agents can interpret results, identify flawed controls and recognize when evidence is insufficient to support a conclusion.
OpenAI evaluated seven of its models across 37 model and reasoning-effort settings, generating 18,500 responses from five attempts at each task.
At the highest shared reasoning setting, GPT-6 Astra earned an average of 66.7% of available grading credit, compared with 62% for GPT-6.1 Sol and 43.8% for GPT-5.6 Sol.
OpenAI said Astra showed improvements in interpreting images, identifying flawed experimental controls, and applying biological context, although challenges remain in accurately counting colonies and drawing conclusions from conflicting evidence.
Arda Göreci, CTO and co-founder of Ligo Biosciences, announced the release on X, “Excited to release BioBasicsBench, a benchmark we developed for @OpenAI to test whether AI agents can interpret real lab data, spot experimental errors, and recognize when the evidence doesn’t support a conclusion.”
He added, “Accelerating biological research starts with getting the everyday work right. BioBasicsBench reveals where models still struggle with routine experimental tasks and help measure whether they’re improving.”
On X, one user pointed out that BioBasicsBench includes tasks that assess whether laboratory evidence actually supports a claim, adding that it is “finally, an AI test where ‘you cannot conclude that’ can be the useful answer.”
Another user described the benchmark as a task-level alternative to evaluating scientific capabilities through general question-answering, particularly for organizations assessing AI agents for laboratory support.
OpenAI cautioned that higher benchmark scores do not necessarily indicate reliable laboratory judgment. The evaluation covered only its own models, leaving their performance relative to competing systems untested.
The company said BioBasicsBench would help measure progress toward AI agents capable of supporting everyday scientific research while identifying shortcomings that must be addressed before broader laboratory use.
