Breaking

GPT-5.4 Proposes Drug Reaction Fix, Validated in 10,000 Lab Runs

June 17, 2026 at 13:12 EDT

Foundation Models
Healthcare
AI Agents

OpenAI said on June 17 that its general-purpose reasoning model GPT-5.4 drove a medicinal chemistry project from literature review through to a validated experimental result, discovering an unexpected way to improve a widely used reaction, according to its announcement. The model was paired with Molecule.one's high-throughput robotic synthesis platform "Maria AI" and a specialized lab, completing roughly 10,000 reactions and human bench-scale verification over about three months.

June 17, 2026 · OpenAI

An AI Chemist Improved a Real Reaction — From Literature to Lab

GPT-5.4 ran a near-autonomous medicinal-chemistry loop — proposing that a mild oxidant like TEMPO could enhance Chan-Lam coupling with primary sulfonamides — then a robotic lab and human chemists validated the result.

10,080

reactions run by the robotic lab across two cycles

~3 mo

from first prompt to shared results (Mar 4 → early June)

88%

of boronic acids saw improved yields (83% of sulfonamides)

The TEMPO effect, measured

Before vs. after the optimized conditions — column height is proportional to each value.

Average yield

16.6%

Before

25.2%

After

Share above 30% yield

15.6%

Before

37.5%

After

In bench validation, 11 of 14 representative reactions improved — and 8 more than doubled in yield.

An end-to-end research loop

Literature & hypothesis

GPT-5.4

→

Experiment design & run

Maria AI robotic lab

→

Analysis & proposals

GPT-5.4

→

Human validation

bench-scale confirm

Why it matters

AI moved from hypothesis to a validated, novel result — broadening substrate scope under mild conditions. External experts called the work novel and practically useful, and one additive was swapped for a cheaper analog (4-hydroxy-TEMPO).

The caveats

Not fully autonomous — human judgment stays in the loop. Generalization to other reactions is untested, and independent reproduction is still needed before broad claims.

Continue reading

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