← Back to The Brief

02

GPT-5 Pro cracks a three-year T-cell mystery in an immunology lab

breakthroughHealthcare

Wednesday, June 24, 2026

Confidence

Medium · — no peer-reviewed publication yet

Evidence

lab blog + corroborating analysis

What happened

OpenAI published a case study describing how immunologist Derya Unutmaz used GPT-5 Pro in late 2025 to revisit a three-year-old puzzle centered on a special type of immune cell that helps the body fight cancer and other illnesses . Unutmaz asked GPT-5 to predict the outcome of an experiment he had already conducted on a CD8+ T cell that targets lymphoma; the model correctly predicted the boost in the cells' ability to kill lymphoma cells, and it could not have gleaned the result from the internet because Unutmaz had not yet published it. OpenAI's writeup is the primary source; secondary analysis at Windows Forum notes Unutmaz works at The Jackson Laboratory for Genomic Medicine.

What it means

This is a credible hypothesis-generation case, not a cure. The mechanism that matters: a domain expert held back data the model had never seen, and the model produced biologically plausible predictions a credentialed scientist could check. That is a real test of generalization, but it is also a test only an expert could run — the value here belongs to labs that can ground-truth the output, not to anyone hoping the model replaces the immunologist. Pharma R&D buyers should read this as a sharpened bar for what "AI co-scientist" pilots have to show.

OpenAI published a case study describing how immunologist Derya Unutmaz used GPT-5 Pro in late 2025 to revisit a three-year-old puzzle centered on a special type of immune cell that helps the body fight cancer and other illnesses . Unutmaz asked GPT-5 to predict the outcome of an experiment he had already conducted on a CD8+ T cell that targets lymphoma; the model correctly predicted the boost in the cells' ability to kill lymphoma cells, and it could not have gleaned the result from the internet because Unutmaz had not yet published it. OpenAI's writeup is the primary source; secondary analysis at Windows Forum notes Unutmaz works at The Jackson Laboratory for Genomic Medicine.

Sources

Apply this today

The hands-on layer the brief points to: a workflow you can run, a tool to test, and today’s 60-second video.

Workflow of the day

Convert Messy Client-Workshop Notes into a Findings-and-Recommendations Skeleton with an Evidence Map

Consulting & Enterprise · tested · saves 2-3 hours per workshop

Full walkthrough →

Tool watch

IBM Research CUGA

CUGA: IBM's agent harness, judged by a developer who actually has to ship one

Read the breakdown →

On this page

What happened
What it means

In this issue

← Back to brief

Stat of the Day

200B

Aampe weekly decisions

The number of individual AI-agent decisions Aampe's platform processes each week across customers including Grab and Swiggy — the operational scale MoEngage just bought. Yahoo Finance/PR Newswire

Today’s Take

Two stories ran on the same day from opposite ends of the agent deployment curve. Anthropic and MoEngage shipped the infrastructure to put agents inside the seats where decisions about employees and customers actually get made. A federal judge in San Francisco told the largest HR software vendor in the country that designing those agents from California means California discrimination law follows the agent wherever it acts. The shape of the day is a widening gap between deployment velocity and accountability velocity — and the buyers who close that gap deliberately, in contracts and audit logs, will own the next eighteen months. The ones who don't will discover that "the vendor did it" is no longer a defense.

— Agentic desk

Never miss a signal

Join professionals getting the daily AI brief every weekday morning.

No spam · Unsubscribe anytime

No sponsored verdicts
Sources linked
Corrections public