OpenAI Releases LifeSciBench -- Frontier Models Pass Just 1 in 3 Real Science Tasks

OpenAI introduced LifeSciBench, an expert-authored benchmark of 750 tasks across seven biological domains, built with 173 scientists. Even the strongest models pass only about one task in three -- a sober counterweight to claims that AI is close to autonomous scientific research.

By the Numbers

750
Tasks
173
Contributing Scientists
~33%
Top Model Pass Rate
7
Domains
Trace Cohen
By Trace Cohen · Value Add Pulse AI Desk
Founder & Editor-in-chief, Value Add VC
1 min read
OpenAI Releases LifeSciBench -- Frontier Models Pass Just 1 in 3 Real Science Tasks: key figures
ShareXLinkedInEmail

THE RUNDOWN

1

A credible ceiling on AI-for-science hype: today's frontier still fails two of three real research tasks

2

Expert-written rubrics from 173 scientists make this a hard benchmark to game -- the gap is real, not artifactual

3

For founders selling 'AI scientist' products, this is the honest baseline buyers will now measure against

The VC Read

Value Add VC analysis

OpenAI publishing a benchmark its own models fail two-thirds of is the most useful thing it shipped this week. The 'AI scientist' pitch has gotten ahead of reality, and now there's a hard, expert-graded number to anchor the conversation. For founders selling autonomous research, this is the baseline your buyers will quote back at you -- so position as augmentation, not replacement, or get caught overclaiming. The honest framing is also the more defensible business.

Analysis

OpenAI released LifeSciBench, a benchmark of 750 expert-authored tasks spanning seven biological domains and seven research workflows, developed with input from 173 scientists. The headline result is humbling: even the most capable frontier models pass only roughly one in three tasks, with detailed rubrics grading the quality of scientific reasoning rather than just final answers.

The benchmark matters because it pushes back against the loudest version of the AI-for-science narrative. While models can accelerate literature review, hypothesis generation, and some analysis, LifeSciBench shows they remain far from autonomously executing real research workflows. The expert-written rubrics make the result hard to dismiss as a measurement artifact.

“The benchmark matters because it pushes back against the loudest version of the AI-for-science narrative.”

For builders, the takeaway is to calibrate. AI is a powerful copilot for scientists, not a replacement, and the honest framing -- augment the researcher, don't replace them -- is both more defensible commercially and more credible with the domain buyers who will increasingly benchmark these claims against tools like LifeSciBench.

ShareXLinkedInEmail

More on

OpenAI →

More from the Pulse archive: Artificial Intelligence coverage → · All topics

Key Sources

1 source
SourceOpenAI

Reported by OpenAI · Analysis by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with The VC Read, a few times a week. Free to subscribe, no spam.

No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗

Read Next

AI
Grok 5 Slips to Q3 -- xAI's 6-Trillion-Parameter Flagship Still in Training
Illustration for: Grok 5 Slips to Q3 -- xAI's 6-Trillion-Parameter Flagship Still in Training
AI

Grok 5 Slips to Q3 -- xAI's 6-Trillion-Parameter Flagship Still in Training

xAI's Grok 5, the rumored ~6-trillion-parameter flagship, has missed both its Q1 and Q2 2026 windows and remains in training, with full API access now looking like a Q3 story. Prediction markets price a public release before June 30 at roughly 7%, even as xAI ships smaller updates like Grok Voice and Grok Imagine Video.

AI
Anthropic Launches Claude Tag, a Persistent AI Teammate That Learns Your Company
Anthropic Launches Claude Tag, a Persistent AI Teammate That Learns Your Company logo
AI

Anthropic Launches Claude Tag, a Persistent AI Teammate That Learns Your Company

Anthropic launched Claude Tag, replacing its old Slack app with a persistent AI teammate that learns a company's context over time, monitors channels, and works autonomously rather than only responding when summoned. It's a bid to move Claude from an on-demand assistant to an always-on coworker embedded in the flow of work.

AI
OpenAI Unveils 'Jalapeño,' Its First Custom AI Chip, Built With Broadcom for Inference at Scale
OpenAI Unveils 'Jalapeño,' Its First Custom AI Chip, Built With Broadcom for Inference at Scale logo
AIFirst custom inference chip

OpenAI Unveils 'Jalapeño,' Its First Custom AI Chip, Built With Broadcom for Inference at Scale

OpenAI revealed Jalapeño, its first in-house silicon -- a chip designed with Broadcom and purpose-built for running AI models (inference) rather than training them. OpenAI says early results show 'significantly better performance-per-watt' than current state-of-the-art alternatives, marking its most concrete step yet to reduce a near-total dependence on Nvidia GPUs.