Claude’s Protein Binder Experiment Points to a New R&D Operating Model

TL;DR

Claude designed protein binders against 14 of 15 targets with interpretable results, confirmed by two external labs. A human expert first encoded the entire process into a ~30,000-token protocol; Claude then coordinated models, compute, and labs autonomously. The results are real but early — performance varied between targets, and Anthropic couldn't isolate whether the model, the pipeline, or run variation drove the differences. Binding is only the beginning; function, safety, dosing, manufacturing, and clinical benefit all still lie ahead. Anthropic's smartest move is staying the intelligence layer — working across thousands of programs without carrying a single asset on its balance sheet.

I didn't want my feed to become an Anthropic news cycle. But a day after I challenged Dario Amodei's cure timeline, Anthropic published results backing the other half of my argument: they are becoming the intelligence layer of drug discovery.

Here's what happened: Claude "designed" wet-lab-confirmed protein binders against 14 of the 15 targets that produced interpretable results.

What 'Claude designed' actually means

Now, what does "Claude designed" actually mean? A human expert encoded the process in a protocol of roughly 30,000 tokens. Two thirds of it dealt with orchestration, verification and operations. The 48-hour campaign ran on up to 12,500 H100 hours of compute.

Claude then installed and validated the available open-source protein models, chose where on each target to bind, assembled a different combination of tools for each target, ran the optimization cycles and selected which designs should be made. Two external labs, Adaptyv Bio and Twist Bioscience, built and tested the proteins.

This is a glimpse of a different R&D operating model: human expertise encoded into a repeatable protocol, with a general-purpose AI coordinating specialist models, compute and the labs that test the output.

How early this still is

The results also show how early this is. Performance varied considerably between targets and campaigns. Opus 4.8 succeeded against TNFα while the generally more capable Mythos Preview failed. Anthropic could not say whether the difference came from the model, the pipeline it chose or ordinary variation between runs.

And obviously, binding is also only the very beginning. Function, safety, dosing, manufacturing and clinical benefit all remain ahead.

Anthropic released the designs and experimental results, including the failures, creating a public benchmark for future protein-design systems.

Why the intelligence layer is the right position

On Tuesday, I wrote: "As the intelligence layer, Anthropic can work on thousands of programs without carrying a single asset on its balance sheet." This is a much smarter position for Anthropic than becoming one more company carrying molecules through the clinic.

Key takeaways

• I see this as a glimpse of a different R&D operating model: human expertise encoded into a repeatable protocol, with a general-purpose AI coordinating specialist models, compute, and external labs.
• The 48-hour campaign required up to 12,500 H100 hours and a ~30,000-token human-authored protocol — two thirds of which covered orchestration, verification, and operations.
• Performance varied considerably between targets, and Anthropic could not attribute the differences to the model, the pipeline, or ordinary run variation.
• Binding is only the first step; function, safety, dosing, manufacturing, and clinical benefit all remain ahead.
• Anthropic released designs and results — including failures — creating a public benchmark for future protein-design systems.
• Working as the intelligence layer lets Anthropic operate across thousands of programs without carrying a single molecule on its balance sheet.
• That is a far smarter position than becoming one more company carrying assets through the clinic.