OpenAI releases LifeSciBench to test AI in real science work
OpenAI released , a test for measuring how useful are in real life science research. It does not just check whether an AI can answer biology facts. It asks to read papers, tables, figures, and data files, then judge evidence, spot problems, suggest experiments, and explain decisions.
has 750 tasks written by experts, with 173 scientist contributors and 453 independent reviewers. Seventy-nine percent of the tasks require several steps of reasoning, and 53% require the AI to use attached materials. OpenAI says improved over GPT-5.5, raising the overall pass rate from 25.7% to 36.1%.
The results still show clear weaknesses when tasks involve heavy use of files, design decisions, exact numbers, sequences, or structures. OpenAI says this test is not the same as proving real-world research impact, so the next step is testing AI inside live reflows.
Key points
- is a benchmark for testing AI on real life science research tasks.
- It includes 750 expert-written tasks, 173 scientist contributors, and 453 independent reviewers.
- Most tasks require multi-step reasoning, and over half require using attached materials.
- reached a 36.1% overall pass rate, compared with 25.7% for GPT-5.5.
- still struggle with file-heavy work, design tasks, and answers that require exact numbers or structures.