Two people at Anthropic understand proteins but have never written code that squeezes a graphics card. In under four weeks, they supervised a model that sped up more than 30 tools widely used by molecular biologists.
That is how the company describes it, and the detail about the two of them is more interesting than the number 4.
Fact
The facts: on 17 September 2026 Anthropic published results in which an internal general-purpose research model, working in Claude Science, optimised more than 30 open models for protein structure prediction and design, genomics and protein language models in just under four weeks. Per the company, the speed-up is about 4 times on average with minimal loss of precision and nearly 2 times with identical outputs. The FlashPairformer kernels, written with Claude, accelerate the core operations in models such as AlphaFold3, OpenFold3 and Boltz-2. A new low-memory mode predicts systems over 10,000 tokens on a single NVIDIA GPU node, among them human mitochondrial complex I and a bacterial ribosome; systems over 31,000 tokens can now be run but are not predicted correctly. The work was supervised by two staff members experienced in biomolecular modelling and without experience in optimisation. Per Anthropic, a single Claude model with one H200 and 24 hours designed protein binders with in silico scores comparable to their earlier campaigns while using about two orders of magnitude fewer GPU hours - about $150 in total for compute and tokens. The code is open. With Adaptyv Bio they announce a protein design competition: up to $1 million in Claude credits, up to $250,000 in Modal credits, DNA from Twist Bioscience and wet lab validation of over 5,000 designs.
The two numbers worth remembering
The first is $150. Their previous protein design campaign allowed up to $10,000 per target. Now a comparable in silico result costs about as much as dinner for four, and if it holds up in the lab, this kind of work moves from big institutes down to small labs.
The second is the honest admission about the giant structures. They can now be run on one machine, but they come out wrong, most likely because the models do not generalise that far beyond what they were trained on. It is good that they say so themselves, with pictures of the failures.
An in silico score is a promise. The lab is the verdict.
That is why the competition matters more than the blog. Five thousand designs will go through a real lab, and there we will see how many of the cheap in silico results become a protein that binds where it should.
The visual is generated code-art, no third-party images.