On September 17, 2026, Anthropic reported that Claude, supervised by two members of its technical staff who had biomolecular modeling experience but no prior experience in inference optimization or kernel engineering, accelerated more than 30 open-source deep learning models for biology in just under four weeks. The set spans structure prediction, protein design, genomics and protein language models, with named examples including AlphaFold3, OpenFold3 and Boltz-2, built on architectures ranging from AlphaFold-class transformers to diffusion, flow matching and graph neural networks. Claude Mythos 5.1, Mythos 5 and Opus 5 were used.
Across more than a dozen structure prediction models, the optimizations gave on average roughly a 4x speed-up with a minimal loss of precision, and roughly 1.6x with identical outputs. Claude's FlashPairformer kernels beat the field's standard implementation by 2.7 to 2.9x on triangle attention and 1.7 to 3.2x on triangle multiplication. It also built a low-memory "Big" mode enabling accurate modeling of systems larger than 10,000 tokens and successful inference on systems larger than 70,000 tokens; on a single 8-GPU B300 node it produced predictions of whole viral capsids and protein compartments of more than 31,000 to more than 70,000 tokens.
The cost result is the most striking. In an earlier protein-design campaign, Claude had been allowed to spend up to $10,000 per target on cloud GPUs. With one NVIDIA H200 and 24 hours of wall time, and the new optimized tools, a single Claude model reached approximately the same in silico quality (ipSAE scores) across 16 targets using about two orders of magnitude fewer GPU hours, for a combined GPU and token spend of about $150. Anthropic open-sourced the optimized code in the uplifting-biomolecular-modeling repository.
The caveats are real. This is a lab reporting on its own model, and the design comparison is in silico against Anthropic's own earlier campaigns rather than wet-lab results. Only the 1.6x figure is output-identical; the 4x figure trades away some precision, and successful inference on very large systems is not the same claim as accurate prediction. What the work does show is the pattern Anthropic is betting on: a small team pointing a model at unglamorous engineering work that makes scientific tooling cheaper for everyone.