Anthropic released a set of code optimizations created with Claude's help for more than 30 open computational biology models. According to the company, the work was completed in less than four weeks and sped up tasks in structure prediction, protein design, genomics, and protein language models by an average of roughly four times, with minimal loss of accuracy. In configurations that preserved identical results, the average improvement was close to two times.

Claude as an optimizer of scientific tools

The announcement shifts the usual focus of artificial intelligence applied to science. Rather than presenting Claude as a replacement for a specialized model, Anthropic used it to modify and accelerate software that researchers already use. The company says an internal general-purpose model worked on dozens of systems, including tools for cofolding, structure prediction, protein generation, chromatin accessibility, and variant assessment.

The public repository makes it possible to examine the scope of the release. It includes kits for projects such as Boltz-2, ColabFold, ESMFold2, RFdiffusion, and Protenix, along with separate instructions for exact, fast, and big modes. Anthropic released the original code under Apache 2.0, although each dependency retains its own license. It also warns that this is a reference release provided “as is,” which will not receive maintenance or contributions and which executes code from external projects.

That clarification matters: the announcement does not amount to independent validation of every optimization or guarantee that all laboratories will obtain the same performance. The figures depend on the model, configuration, hardware, and type of computation. The repository pins versions, commits, and digests for dependencies to facilitate reproduction, but users still need to review each project's licenses, vulnerabilities, and computing requirements.

Speed, memory, and large systems

Anthropic reports an approximate average speedup of four times with minimal changes in accuracy and roughly 1.6 times across the broader set shown in its charts; for runs with identical outputs, it summarizes an improvement of close to two times. In specific triangle attention and triangle multiplication operations, it claims to outperform a reference benchmark by between 1.7 and 3.2 times, depending on the configuration. These are company measurements and should be read as technical results of the release, not as a universal comparison.

The big mode addresses another bottleneck: memory. Anthropic says it allows accurate modeling of systems with more than 10,000 biomolecular “tokens”—amino acids, nucleotides, small-molecule atoms, and ions—on a single NVIDIA GPU node. The official image shows complexes such as the 70S ribosome and human complex I with more than 10,000 tokens.

The company also managed to run inference on systems with more than 70,000 tokens on a single B300 node, but acknowledges a decisive limitation: the predicted structures collapsed. Those examples were almost two orders of magnitude beyond the models' training context. The result therefore demonstrates that the computational barrier to attempting calculations at that scale was lowered, not that those systems can be modeled correctly today.

Open-source code and experimental validation

The release is accompanied by a protein-design competition organized with Adaptyv Bio on Proteinbase. The program poses five problems, including cross-species reactivity, pH sensitivity, peptide-MHC specificity, and GPCR targets. Anthropic is offering up to one million dollars in Claude credits; Adaptyv and its collaborators plan laboratory validation for more than 5,000 designs.

The combination matters because it connects three stages that are usually evaluated separately: modification of scientific software, execution of biomolecular models, and experimental verification. The open-source code will allow researchers to audit whether the improvements hold up outside Anthropic's environment. The competition, for its part, may provide physical evidence about designs produced with those tools. Until external reproductions and laboratory results exist, the speed figures should be treated as results reported by Anthropic, even though the code is already available to examine them.