On September 17, Anthropic published a research report describing how it had Claude optimize inference for more than 30 open-source biomolecular models, cutting their runtime by about 4x on average in under four weeks, with only a small loss in accuracy. When accuracy loss isn't allowed at all, the average speedup is about 1.6x. All the code is now on GitHub, in the anthropics/uplifting-biomolecular-modeling repository.
The models span four categories: structure prediction, protein design, protein language models, and genomics models. The report names AlphaFold3, OpenFold3, Boltz-2, and ColabFold 1.6.3; the design category covers hallucination-based design, structure generation, and inverse folding, built on architectures ranging from diffusion to flow matching to graph neural networks. The report doesn't list the remaining models individually.
Two people and one model
The project was overseen by two Anthropic engineers. The report specifically notes their background: they knew biomolecular modeling well, but had never done inference optimization or written a GPU kernel before. The actual work was carried out by Claude inside the Claude Science environment; the model versions tested were Mythos 5.1, Mythos 5, and Opus 5.
The speedups come in two layers. One is generic: Claude wrote a custom kernel called FlashPairformer specifically for the Pairformer module, the most compute-heavy part of these models, speeding up triangle attention by 2.7x to 2.9x and triangle multiplication by 1.7x to 3.2x. The other layer is model-specific — things like caching repeated computations and pruning branches that never get used.
The report also spells out how it judges "only a small loss" in accuracy: for interface prediction, it uses the DockQ score, with anything above 0.23 counted as acceptable; for design results, it uses ipSAE, a computed score that correlates with wet-lab binding rates.
How big a structure it can fold
Beyond speed, the more striking change is in scale. AlphaFold3's previous prediction ceiling was 7,663 tokens; after optimization, it now runs past 70,000 tokens on a single NVIDIA B300 node, large enough for targets like entire viral capsids and protein compartments.
"These are among the largest structures ever folded accurately using structure prediction models."
The report is upfront about the limits: predictions above 70,000 tokens aren't accurate. The hardware bar is also real — large assemblies need a B300 node, and even routine jobs still need an H200-class GPU.
How it lines up with last month's design round
In mid-August, Anthropic published an experiment in which Claude designed protein binders from scratch, getting usable designs for 14 of 15 targets — this site covered it at the time. That round used multiple subagents and heavy prompting running on Modal's cloud, and according to this new report, cost about $10,000 per target.
This time, Anthropic redid protein design with the optimized model stack. Across a test of 16 targets, a single H200 running for 24 hours of wall-clock time brought combined GPU and token costs to about $150, which Anthropic says is roughly two orders of magnitude cheaper. The targets in the two rounds aren't identical, and the report doesn't give wet-lab validation data for this round's designs, so cost and turnaround time are the only things directly comparable for now — hit rate will have to wait for further results.
From August to September, Anthropic's sequence in life sciences is clear: first show the model can design molecules that hold up in the lab, then go back and rebuild the design tooling itself. That second step may matter more directly for academic labs — many groups have only a handful of GPUs, and large structures that were previously out of reach now have an open-source, accelerated version they can try.
Sources: Anthropic research report, CocoLoop, the anthropics/uplifting-biomolecular-modeling code repository; check the original report for the exact average speedup multiple, FlashPairformer's per-module speedup ranges, the 7,663- and 70,000-token limits, and the per-target cost figures.