Breaking

OpenAI Unveils LifeSciBench, a Life-Science AI Benchmark Built With 173 Scientists

June 17, 2026 at 16:42 EDT

Research & Papers
Foundation Models
Healthcare

On June 17, 2026, OpenAI introduced LifeSciBench, a benchmark for measuring how well AI supports real-world life science research. Developed jointly with 173 scientists from biotechnology and pharmaceutical research, it comprises 750 expert-authored tasks spanning seven biological research workflows.

June 17, 2026 · OpenAI

LifeSciBench: Grading AI on Real Lab Science

A new benchmark built with 173 biotech and pharma scientists puts 750 expert-authored tasks across seven research workflows — from drug discovery to genomics — to measure whether AI can handle the messy judgment calls of real-world research.

750

expert-authored tasks across 7 workflows

173

Ph.D.-level task creators

453

expert reviewers

19,020

rubric criteria (~25 per task)

GPT-Rosalind vs GPT-5.5 on LifeSciBench-related benchmarks

Scores shown to scale (higher is better). GPT-Rosalind leads each.

27.5%

25.1%

MedChemBench

21.6%

20.4%

GeneBench

63.2%

55.8%

LabWorkBench

GPT-Rosalind GPT-5.5 + token efficiency gains of −5.3% to −31%

79%

of tasks are multi-step (~4 steps each)

53%

require interpreting artifacts

1,062

artifacts: figures, PDFs, sequence & chemical files

How tasks are scored

Expert-authored rubrics grade more than the right answer — they judge the reasoning behind it.

Correctness Scientific validity Practical usefulness Justification Caveats Formatting

WELCOMED

An expert-reviewed benchmark that holds AI accountable for the quality of its scientific decisions — not just factual recall.

STILL EARLY

New benchmark with only a partial third-party leaderboard release; broad, long-term user case studies have yet to accumulate.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.