What is a protein binder?

A protein binder is a small, designed protein that attaches tightly to a specific target protein, similar to how an antibody locks onto its target. Anthropic tested the binders for binding strength in the lab, not for safety or effectiveness as medicines.

Can outside researchers use Claude to design protein binders themselves?

Not broadly yet — Anthropic says protein-design capability remains restricted and isn’t available for general access in Claude Fable 5, citing dual-use biosecurity concerns.

Does this mean AI just designed a new drug?

No. Anthropic stresses a protein binder is only a first step toward a therapeutic — turning it into an actual drug-like molecule takes considerably more development.

Anthropic Says Claude AI Designed Working Protein Binders in Lab Tests

Q: Which Claude models were used, and how were the results checked?

Claude Mythos Preview and Opus 4.8 generated the designs; independent labs Adaptyv Bio and Twist Bioscience built and tested them in the wet lab to confirm which ones actually bound their targets.

Anthropic says its Claude AI models designed protein binders — molecules that attach tightly to a target protein — that worked against 14 of 15 disease-relevant targets in independent wet-lab testing, according to research the company published on August 18.

How the test worked

Two Claude models, Mythos Preview and Opus 4.8, generated 1,320 candidate protein designs across 15 targets with no human guidance during the design process itself, Anthropic said. Independent labs Adaptyv Bio and Twist Bioscience built and tested the molecules, confirming 354 working binders.

Working alone against all 15 targets over 48 hours, Mythos Preview hit a 26.7% success rate and Opus 4.8 hit 22.6%, according to the company. When Mythos Preview instead spent 24 hours focused on one target at a time, its hit rate rose to 35.1% — roughly double to triple the 10% to 15% rate Anthropic says is typical across protein-design campaigns today. Results varied widely by target, from zero confirmed binders on maltose-binding protein to a 90% hit rate on the easiest one.

The starkest gap came on a target called RBX1: Mythos Preview reached a 40% hit rate, compared with 3.7% among human entrants in an Adaptyv Bio design competition, and its top design reportedly outperformed the winning entry out of 245 submissions.

Not a drug — yet

Anthropic was careful to frame the results as an early step, not a finished therapy. “Protein minibinders are not a standard therapeutic modality,” the company said, and turning a binder into a drug-like molecule takes substantially more work. The biology capability itself also remains restricted: Anthropic has not opened general protein-design access in Claude Fable 5, its most capable model, citing dual-use risks tied to biological research — a caution it has extended to other systems, including its Mythos 5 access limits.

The company also reported that a separate Claude model processed lab spectroscopy data — NMR and mass-spectrometry readings — in under 25 minutes with accuracy close to human lab analysis, part of a broader push to apply Claude across experimental chemistry alongside drug-discovery workflows.

How the test worked

Not a drug — yet

Frequently asked questions