Claude rewrote the code behind 30+ biomolecular models in four weeks — and made them ~4x faster on average

On September 17, Anthropic published a technical report showing that Claude optimized more than 30 open-source biomolecular models (36 implementations) in under four weeks, spanning structure prediction, protein design, protein language modeling and genomics. The supervisors were two researchers experienced in biomolecular modeling but with no prior background in inference optimization or GPU kernel engineering. The notable part is what AI was used for: not interpreting experimental data, but performance engineering on the software scientists actually run.

The numbers come in layers. First, speed: on NVIDIA H100s, Exact mode, which reproduces the unmodified model's outputs bit for bit, ran the forward pass of 14 structure-prediction models 1.6x faster on average; Fast mode, which trades a small amount of precision for speed, ran 13 models 4.1x faster on average. Individual results varied widely — Exact gains ranged from 1.45x to 7.12x and Fast gains from 2.91x to 31.07x — so four times is an average, not a promise. Second, memory: Claude built a low-memory Big mode that lets a single 8-GPU node accurately predict molecular machines larger than 10,000 tokens — human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial ribosome — where multi-node inference was previously required. Third, kernels: Claude produced reusable FlashPairformer v1 GPU kernels whose triangle attention runs on average 2.7x faster than the field-standard implementations.

The report is candid about limits. Asked to push further, Claude ran inference on complete viral capsids and protein compartments from 31,000 to over 70,000 tokens on a single 8-GPU B300 node — inference succeeded, but the structures were not predicted correctly. A program that runs is not the same as a result that is scientifically right; the models do not generalize nearly two orders of magnitude beyond their training context.

Quality checks matter here. Across 13 model configurations and 1,925 model-target pairs, the share of acceptably predicted interfaces (DockQ at least 0.23) changed by less than one percentage point in every mode, with no pooled difference distinguishable from zero. These benchmarks were run inside Anthropic's project rather than by an independent evaluator, which is worth stating plainly.

The economics experiment at the end is the most interesting part. In an earlier binder-design study, Claude could spend up to $10,000 per target on cloud GPUs, roughly 2,500 H100 hours. The new setup gave a single Claude model one NVIDIA H200 for 24 hours, a prompt of about 1,100 words, no subagents and no human steering; across 16 targets, in silico scores approached those of campaigns using about 100 times the GPU budget. Anthropic has open-sourced all optimized code and co-sponsored a protein-design competition with Adaptyv Bio: more than 5,000 designs will be synthesized and tested in wet labs, backed by $1 million in Claude credits. The real test will happen there — whether faster tools actually produce better proteins.

What makes the project a turning point rather than a benchmark stunt is the division of labor it implies. The researcher still chose the models, defined acceptable behavior, inspected outputs and verified downstream performance; Claude handled the engineering that previously required a separate specialization in GPU memory movement and numerical precision. The report does not claim that expertise can be removed from the process. It claims something narrower and more useful: an experienced scientist can now use an AI coding system to take on performance work that would otherwise sit out of reach. Whether that holds at scale, and whether the community can convert these one-off optimizations into durable, retrainable knowledge, is the open question the field will be testing for the rest of the year.