Meet LifeSciBench: grading AI on real life-science research!

Hey everyone, it's Shii! Today I'm excited about a benchmark that measures how far AI can go in helping with real life science research.

OpenAI News openai.com

What was announced?

OpenAI's News just introduced LifeSciBench, a new benchmark. In short, it's a test suite for measuring how well AI can handle the actual research work of drug discovery and biology.

Most biology benchmarks lean on textbook-style Q&A. But real researchers interpret ambiguous evidence, reconcile conflicting results, design hard experiments, and decide what to do next under uncertainty. LifeSciBench is built to measure exactly that messier, multi-step work.

Why it matters

AI agents keep getting better at scientific tasks, but "genuinely useful to researchers" is a separate question. Older evaluations tend to favor narrow domains and clean reference answers, so they miss the broad span of research-level work.

What makes LifeSciBench special is that every task is grounded in the judgment of 173 Ph.D.-level working scientists with biotech and pharma experience. It tries to measure "can AI help with science?" through the eyes of practicing experts.

What changes

Now we can compare models' "life science ability" on a shared yardstick, not just knowledge quizzes but the whole research flow: handling evidence, analysis, experimental design, and how conclusions are communicated.

OpenAI also opened a contributor sign-up for people who want to help build benchmarks. The research world and AI evaluation are getting a lot closer.

Dive Deep

LifeSciBench has 750 tasks across seven workflows and seven biological domains. The workflows are: evidence handling, analysis, design/optimization/prediction, scientific reasoning, validation and operations, translation, and scientific communication.

Tasks aren't multiple choice; they're free-response, written like a request to a knowledgeable colleague. About 79 % need multi-step reasoning, averaging four steps each. They come with 1,062 attached artifacts (figures, PDFs, tables, sequence files, structure/chemical files, web references), and more than half (53 %) require reading those artifacts to solve.

Grading is strict. Expert-written rubrics hold 19,020 criteria total (about 25 per task), awarding points for a specific fact, a reasoning step, or a numeric answer within tolerance. Quality control is layered: accepted tasks passed an average of six automated review cycles and at least two expert reviews, and 453 reviewers (97 % with doctorates) validated their quality.

So how did models do? The life-science-tuned GPT-Rosalind led with an exact pass rate of 36.1 %, up from GPT-5.5's 25.7 % (and ahead of the field including Gemini 3.1 Pro and Grok 4.3). Its strongest areas were scientific communication (56.3 %→71.1 %) and translation (36.8 %→57.7 %) — organizing and communicating evidence is improving fast.

But the weak spots are clear too. Design/optimization/prediction sat at 30.7 % and analysis at 30.3 %. Artifact-heavy tasks were the biggest gap: GPT-Rosalind scored 45.1 % on text-only tasks but dropped to 28.1 % once figures or URLs were involved. Exact outputs were hard as well, with 14.8 % on numeric tasks and 24.0 % on sequence/structure outputs.

OpenAI is careful to note that a high LifeSciBench score doesn't guarantee real-world discovery. It measures task-level capability, and the next step is longer-horizon studies in live research settings. For the full methods and numbers, check the preprint paper (PDF).

Wrap-up

• LifeSciBench is a 750-task benchmark for how well AI supports real life science research
• Built by 173 Ph.D. scientists, with 19,020 criteria and 453 reviewers backing quality
• 7 workflows x 7 domains, 1,062 artifacts; 79 % are genuinely multi-step tasks
• Even top-ranked GPT-Rosalind hits only 36.1 % exact pass rate - strong at communicating evidence, weak at reading artifacts and exact outputs
• OpenAI stresses a high score signals capability, not guaranteed discovery

If you're a drug-discovery or bio researcher, or you care about evaluating science-helping AI, this one's right up your alley! Curious folks can also check the GPT-Rosalind access request.