Anthropic has published one of the more practical demonstrations yet of a general-purpose frontier model improving scientific infrastructure rather than simply producing scientific answers. In less than four weeks, Claude optimized more than 30 open-source models used for protein-structure prediction, protein design, genomics and protein language modeling.
Anthropic says the optimized workloads ran roughly four times faster on average in modes that allow small documented numerical differences. In Exact modes, where outputs are intended to match the original implementations, the average gain was closer to twofold. The resulting code is now public as 36 separate optimization kits covering a broad collection of research tools.
Claude did not build a new biology model — it optimized how existing ones run
That distinction matters. This is not a new competitor to AlphaFold3 or Boltz-2. Claude did not train a new structure-prediction network. Instead, it acted as a performance engineer: inspecting existing implementations, identifying computational bottlenecks and rewriting portions of the software to make better use of GPUs.
The public repository contains kits for structure prediction and co-folding, binder and sequence design, inverse folding, genomics and protein language models. Supported projects include Boltz-2, OpenFold3, ColabFold, RFdiffusion, ProteinMPNN, Evo 2 and a number of other research tools.
Why it matters: making an existing scientific model several times faster can be almost as valuable as inventing a better model. If researchers can obtain the same result with less compute, or fit a workload onto one node instead of several, they can test more hypotheses with the same budget.
FlashPairformer targets one of the most expensive parts of modern structure models
Models in the AlphaFold3, OpenFold3 and Boltz-2 family spend a significant fraction of their runtime and memory on triangle attention and triangle multiplication, operations used to model geometric relationships in complex biomolecular systems.
The difficulty is scaling. Anthropic notes that these operations have cubic cost: doubling the size of a system can require roughly eight times as much compute and memory, while tripling it can require about 27 times as much.
Working with Claude, Anthropic developed a set of custom GPU kernels called FlashPairformer. In the company’s benchmarks, those kernels outperformed its field-standard comparison by about 2.7–2.9× for triangle attention and roughly 1.7–3.2× for triangle multiplication, depending on configuration.
Claude also performed model-specific optimization. Anthropic describes techniques including caching intermediates that the original code recomputed, simplifying branches whose outputs were effectively constant, fusing GPU operations and applying other conventional but labor-intensive performance-engineering techniques.
The memory result may be more important than the speed result
For several structure models, Claude produced a separate Big mode optimized to reduce peak GPU-memory use rather than maximize throughput.
Anthropic says Big mode enabled accurate prediction of biomolecular systems containing more than 10,000 tokens on a single NVIDIA GPU node. In this context, tokens can represent amino acids, nucleotides, atoms from small molecules and ions.
The successfully modeled systems included human mitochondrial complex I, the TRiC chaperonin, a proteasome and a bacterial 70S ribosome. Anthropic says the resulting predictions closely matched experimentally determined structures. For context, it points to the AlphaFold3 paper’s accurate prediction of a 40S ribosome containing 7,663 tokens.
That does not mean arbitrarily large molecular assemblies can suddenly be predicted correctly. Anthropic deliberately pushed the system further, running structures from roughly 31,000 to more than 70,000 tokens — including entire viral capsids and protein compartments — on a single eight-GPU B300 node. The computation completed, but the predicted structures collapsed into incorrect compact forms.
The software ceiling moved; the model’s biological knowledge did not
That failure is one of the most useful parts of the result. Better memory efficiency can let a trained network process an input that was previously computationally impossible, but it cannot teach that network biological relationships it never learned during training.
If an assembly is tens of times larger than the model’s training context, enough GPU memory does not guarantee a correct structure. Anthropic explicitly reports that boundary rather than treating successful execution as successful scientific prediction.
Protein design went from thousands of GPU-hours to one H200
Anthropic also used the optimized software to revisit a previous protein-design experiment.
In the earlier campaign, Claude was allowed to orchestrate sub-agents and spend as much as $10,000 of infrastructure budget per target, equivalent to roughly 2,500 NVIDIA H100 GPU-hours. In the new setup, one Claude model received access to a single H200 for 24 hours, a much shorter prompt and the newly optimized biomolecular tools, without sub-agents or human steering of the design process.
Across 16 targets, Anthropic says three Claude models produced in-silico binding scores comparable to the earlier campaign while using about two orders of magnitude fewer GPU-hours. Including both GPU and model-token costs, the company estimates a run at roughly $150.
The qualification is important: these are computational binding metrics, not proof that a protein works in a living system, and certainly not evidence of a successful drug. Physical synthesis and wet-lab testing are still required.
That is why more than 5,000 designs are going into the lab
Anthropic and Adaptyv Bio are using the release to launch a public Protein Design Competition covering five challenges tied to biological mechanisms and potential drug-design problems.
The two organizations are jointly funding $1 million in experimental validation for more than 5,000 protein designs in Adaptyv’s automated laboratory. Anthropic is also offering up to $1 million in Claude credits; Modal is contributing compute credits, while Twist Bioscience is providing DNA for the experiments. Proteinbase says experimental results for tested designs — including negative results — will be published openly.
The five competition rounds are scheduled from September 28 through October 31. Selected designs are planned for experimental validation by the end of November, with results scheduled for publication on December 15.
Why this is more interesting than another “AI wrote code” demo
GPU optimization is unusually measurable. A rewritten kernel either produces the required scientific output faster or it does not; a memory optimization either lets a larger system fit onto available hardware or it does not.
That makes this a relatively clean example of a general-purpose AI model improving specialized scientific tooling without itself needing to become the world’s best biology model.
The scale is also notable. Anthropic says the 30-plus models were optimized in under four weeks under the supervision of two technical staff members who understood biomolecular modeling but had no prior expertise in inference optimization or GPU-kernel engineering.
The results should not be generalized to every accelerator or workload. Performance depends on the model, operating mode and GPU, and some faster modes trade additional memory for speed. Anthropic also labels the GitHub project a reference release: the company does not plan to maintain all 36 kits as a continuously supported upstream project.
Still, the broader shift is significant. In this project Claude was not merely summarizing papers or proposing molecules. It was changing the computational substrate used by researchers — making existing scientific models cheaper to run, able to tackle larger inputs and easier to use in experimental workflows.
