Breaking

OpenAI Unveils LifeSciBench to Score AI on Life-Science Research

June 17, 2026 at 19:01 EDT

Research & Papers
Foundation Models
Healthcare

OpenAI on June 17, 2026 introduced LifeSciBench, a new benchmark for evaluating real-world life-science research tasks. A total of 173 scientists from biotech and pharma helped build it, contributing 750 expert-authored tasks.

June 3, 2026 · OpenAI

LifeSciBench: testing AI across real life-science research

A new benchmark built with 173 scientists measures AI across 750 real-world life-science tasks — and grades GPT-Rosalind on the full end-to-end research workflow, not isolated trivia.

173

Biotech & pharma scientists who built it

750

Real-world research tasks evaluated

50+

Scientific databases linked via the Codex plugin

The end-to-end research workflow

Prior benchmarks scored isolated knowledge. LifeSciBench grades a full research sequence.

Evidence handling

→

Analysis & design

→

Scientific reasoning

→

Validation & comms

GPT-Rosalind, evaluated head-to-head

A purpose-built model for biology and drug discovery — said to lead on overall and workflow-specific scores against rival models.

GPT-Rosalind

Grok 4.3

Gemini 3.1 Pro

others

Research preview · enterprise trusted-access program with eligibility review · pricing & API undisclosed

Why researchers like it

It tests whether a model knows when it lacks enough evidence to act — reasoning from data and handling uncertainty, not just recalling trivia.

The caveat

Large-scale real-world usage reports remain limited so soon after release; practical assessments are expected to grow over time.

Continue reading

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