Claude accelerates AI models for protein design

Anthropic says Claude optimized more than 30 biology models in less than four weeks and made them run around four times faster on average, while using much less memory. The goal is concrete: give more researchers the ability to predict protein structures and design molecules without needing large computing centers.

The company has published the optimized code and announced, together with Adaptyv Bio, a protein design competition offering up to $1 million in Claude credits, $250,000 in computing resources, and experimental validation for more than 5,000 designs.

What Claude optimized

Biology models can predict a protein’s three-dimensional shape from its amino acid sequence. They can also design new proteins to bind to a specific molecule, block a signal, or transport another substance.

These tools are used in drug research, genetics, and disease studies. The problem is that some calculations grow rapidly as the system gains more components. In models such as AlphaFold3, OpenFold3, and Boltz-2, certain operations can require 8 times more time and memory when doubling the system’s size.

Claude created specific optimizations for those calculations, including new software components for graphics cards called kernels. According to Anthropic, the result was:

• Between 2.7 and 2.9 times more speed in one of the main operations.
• Between 1.7 and 3.2 times more speed in another, depending on the configuration.
• An average acceleration of around 4 times in structure prediction models.
• Nearly 2 times more speed without changing the output results.

Anthropic says the accelerated models maintained their performance on the biological tasks evaluated. The company also notes that this type of optimization usually takes weeks of work from specialized engineers for each model, while Claude worked under the supervision of two team members with experience in biomolecular modeling but no prior experience in inference optimization or kernel development.

Large molecules on a single machine

The reduction in memory use makes it possible to work with larger biomolecular systems. Claude created a mode called Big that, according to Anthropic, can accurately predict systems of more than 10,000 tokens on a single node with an NVIDIA GPU. In this context, a token can represent amino acids, nucleotides, atoms from small molecules, or ions.

The modeled systems included human mitochondrial complex I, the TRiC chaperonin complex, a proteasome, and a bacterial ribosome. According to the company, each came close to its experimentally determined structure.

The mode also made it possible to run predictions for systems of more than 70,000 tokens on a single node. That does not mean all of those predictions are correct. Anthropic acknowledges that the models failed to accurately predict viral capsids and protein compartments ranging from 31,000 to 70,000 tokens, but they were able to process them, which had previously been much more difficult.

Designing proteins with fewer resources

Anthropic tested its system again to create proteins that bind to specific targets. This time it used a single Claude model, an H200 GPU, 24 hours, and a prompt of around 1,100 words, with no subagents or human intervention during the design process.

Across 16 targets, the Mythos 5.1, Mythos 5, and Opus 5 models reached design scores similar to those from an earlier campaign, but using approximately 100 times fewer GPU hours. The company estimates that the combined cost of GPUs and tokens was around $150.

This is a computational comparison, not proof that the proteins work in patients or even in a laboratory. To verify that, they need to be synthesized and their real-world behavior measured.

What changes for researchers

Open-source code can reduce the cost of experimenting with these models and allow smaller labs to work with systems that previously required several GPUs or large budgets. That may speed up the exploration of drug candidates, but it does not eliminate the experimental work or guarantee medical results.

The competition with Adaptyv Bio will test problems such as how a protein reacts across different species, its sensitivity to pH, its specificity toward peptide-MHC complexes, and its design against GPCR receptors. The selected designs will be evaluated in a laboratory.

The important signal is not that Claude has solved protein design. It is that AI is starting to reduce one of science’s less visible bottlenecks: the time and cost required to make the tools work. The next step will be to see how many of those designs survive laboratory testing, where predictions stop being promises and become results.