Claude optimised 30+ biomolecular models in four weeks

Anthropic has handed protein scientists a large discount on their electricity bill. In a research post published on 17 September, the company says Claude optimised more than 30 open-source biomolecular models in just under four weeks, making them roughly four times faster on average and meaningfully lighter on memory. The code is being open-sourced, and a protein design competition with Adaptyv Bio is open for entries.

The context matters as much as the result. Anthropic's previous protein design demonstration had Claude orchestrate sub-agents with a 16,000-word prompt and spend up to $10,000 per target on Modal, roughly 2,500 NVIDIA H100 GPU hours. Impressive, and completely out of reach for most protein designers. This time the same class of result came from a single Claude model given one NVIDIA H200, 24 hours, a prompt of about 1,100 words and no human steering — with a combined GPU and token spend of roughly $150.

What Claude actually optimised

Structure prediction models such as AlphaFold3, OpenFold3 and Boltz-2 burn most of their runtime and memory on two operations: triangle attention and triangle multiplication, which are cubic in both. Double the size of the system and they need eight times the time and memory; triple it and you are looking at 27 times. Working with Claude, Anthropic wrote FlashPairformer, a set of custom GPU kernels that beats the field standard by 2.7 to 2.9 times on triangle attention and 1.7 to 3.2 times on triangle multiplication, depending on the model.

Claude then went model by model, doing the unglamorous engineering: caching work that was being recomputed, collapsing dead branches into their constant outputs, and generally tightening inference paths. Across more than a dozen structure prediction models that produced about a 4x speed-up with minimal loss of precision, and about 1.6x even when the outputs had to be identical. Protein design models built on diffusion, flow matching and graph neural networks got the same treatment, as did genomics and protein language models.

What makes the exercise notable is who was holding the spanner. Claude was supervised by two members of Anthropic's technical staff with biomolecular modelling backgrounds but no prior experience in inference optimisation or kernel engineering — work that normally takes an experienced team weeks per model, with optimisations that often do not transfer between models.

Big mode and a genuinely honest caveat

The memory work is the part biologists will feel daily. A new low-memory "Big" mode accurately models systems larger than 10,000 tokens and runs inference on systems beyond 70,000 tokens on a single GPU node, something that previously meant spreading work across a cluster. Anthropic says it folded human mitochondrial complex I, the TRiC chaperone complex, a proteasome and a bacterial ribosome, each closely matching its experimentally determined structure.

Credit where it is due: the same post admits that the capability runs were proof-of-concept using a single trunk pass, and that pushing nearly two orders of magnitude beyond the training context makes predicted structures collapse. That is a more useful sentence than a benchmark table, because it tells researchers exactly where the floor gives way.

The competition, and what is still missing

Anthropic and Adaptyv Bio have picked five problems at the edge of current protein design — species cross-reactivity, pH sensitivity, peptide-MHC specificity and difficult targets such as GPCRs — and will put more than 5,000 community designs through wet-lab validation. There is up to $1 million in Claude credits on the table, up to $250,000 in Modal compute credits and DNA supplied by Twist Bioscience. Anthropic also opened its Life Sciences Verification Program into public beta, with its first organisations enrolled.

Our opinion

This is the most interesting flavour of AI announcement: the model was pointed at expert software engineering rather than at writing, and it beat the humans on volume without being handed a specialist toolchain. Kernel work is about as far from a content prompt as you can get, and the fact that two biologists with no GPU-optimisation background could direct it is the detail that should worry anyone whose job is tuning inference code by hand.

The number to remember is not the 4x, though. It is the $150. Turning a $10,000-per-target experiment into pocket change is what changes who gets to do this research at all — and it makes the competition's wet-lab results the thing worth watching, because in-silico binding scores have a long history of flattering themselves.