The Pulse

Anthropic Open-Sources Claude’s Biology Model Optimizations

Claude optimized more than 30 open-source models used in biomolecular research, speeding them up roughly fourfold on average, Anthropic says, while reducing memory use for large molecular systems.

AI.info Team · Written with AI · fact-checked · 2026-09-18

Anthropic’s Claude has optimized more than 30 open-source models used for biomolecular structure prediction, protein design, genomics, and protein language modeling. Claude completed the work in just under four weeks under the supervision of two Anthropic technical staff members experienced in biomolecular modeling.

The optimizations produce an average speedup of roughly four times with minimal loss of precision, and nearly two times faster execution when outputs must remain identical. The release is aimed at researchers who depend on open biological software but may not have access to the large GPU clusters required by many current workflows.

Open-source optimizations, not new biology models

Anthropic is not releasing a new Claude model or replacing the underlying biological models. Instead, it is open-sourcing the optimized code for more than 30 open-source models spanning structure prediction, protein design, protein language modeling, and genomics.

We are open-sourcing the optimized code for all of these models today so that the broader community can make use of them.

Anthropic research team

The code is intended to make existing biological modeling workflows faster and more accessible. The work covers models trained for tasks including structure prediction and protein design, as well as genomics and protein language modeling.

FlashPairformer targets the expensive part of structure prediction

Much of the runtime in modern structure-prediction systems such as AlphaFold3, OpenFold3, and Boltz-2 goes into triangle attention and triangle multiplication. Those operations model relationships among triplets of tokens, but their time and memory requirements grow cubically as the molecular system expands.

Claude helped develop a set of custom GPU kernels called FlashPairformer for the two operations. FlashPairformer outperformed the field standard by an average of 2.7 to 2.9 times for triangle attention and 1.7 to 3.2 times for triangle multiplication, depending on the model configuration, Anthropic reports.

Claude also produced model-specific changes, including caching redundant recomputed work and simplifying dead branches into their constant outputs. The resulting structure-prediction models ran four times faster on average, Anthropic says, with no measured effect on the downstream task. Across more than a dozen structure-prediction models, the company reports roughly 1.6 times faster execution when outputs were identical and roughly four times faster execution with only a minimal precision reduction.

Big mode brings larger molecular machines into range

The release also addresses memory use. Anthropic’s low-memory “Big” mode successfully modeled systems larger than 10,000 tokens and ran inference on systems larger than 70,000 tokens using one NVIDIA GPU node.

Successful examples included human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial ribosome. Each closely matched its experimentally determined structure, Anthropic says. Complex I and the 70S ribosome each contain more than 10,000 tokens, compared with 7,663 tokens for the 40S ribosome predicted accurately by AlphaFold3.

The system can run larger proof-of-concept jobs as well. Using one eight-GPU NVIDIA B300 node, Anthropic generated predictions for viral capsids and protein compartments ranging from more than 31,000 to more than 70,000 tokens. Those predictions were not correct, but they ran at a scale that previously required a higher computational barrier.

Protein design costs fall from thousands of dollars to about $150

Anthropic used the optimized models in a second protein-design experiment. Earlier work gave Claude an approximately 16,000-word prompt, access to sub-agents, and up to $10,000 per target on Modal, equivalent to about 2,500 NVIDIA H100 GPU hours.

For the new test, one Claude model received access to a single NVIDIA H200 for 24 hours, a prompt of about 1,100 words, and a reference sheet for the pre-installed tools. Anthropic ran Mythos 5.1, Mythos 5, and Opus 5 against 16 targets and evaluated the resulting designs with ipSAE, an in-silico score used to estimate binding performance.

The median- and highest-scoring designs reached approximately the same ipSAE values as the earlier Mythos 5.1 campaigns while using about two orders of magnitude fewer GPU hours, Anthropic says. The company estimates that combining GPU and token costs brought the spend to approximately $150 for comparable in-silico performance.

Adaptyv Bio competition will test more than 5,000 designs

Anthropic is pairing the code release with a protein-design competition co-sponsored by Adaptyv Bio. The organizers selected five problems, including species cross-reactivity, pH sensitivity, peptide-MHC specificity, and difficult targets such as G protein-coupled receptors.

Adaptyv will experimentally validate more than 5,000 community-submitted designs. Anthropic is offering up to $1 million in Claude credits, while Modal is providing up to $250,000 in compute credits. Twist Bioscience will supply DNA for the competition.

The distinction between prediction and laboratory validation matters. Anthropic’s speed and cost figures come from computational evaluations, while the competition is intended to test how many designs survive experimental screening. The open-source release lowers the computational cost of existing workflows; it does not by itself establish that the resulting molecules will work in a lab.

Anthropic has also opened a public beta of its Life Sciences Verification Program. The company has enrolled its first group of organizations to provide frontier AI capabilities for biology-related work.