Anthropic published a research paper on September 17, 2026, reporting that Claude optimized over 30 open-source biomolecular models in just under four weeks, which the company said accelerated the models by about 4x on average. Anthropic open-sourced the optimized code and announced a protein design competition co-sponsored by Adaptyv Bio .
36 optimized packages and FlashPairformer cores
According to the research paper and an accompanying technical report , an internal general-purpose research model generated optimized packages for 36 model
implementations, covering over 30 open-source models in six families: co-folding and structure prediction (14 packages), hallucination (3), structure generation (6), reverse folding (3), genomics (7), and protein language models (3). Anthropic said that Claude was supervised primarily by two members of its technical staff with experience in biomolecular modeling, but no experience in inference optimization or kernel engineering.
Each package runs with the original model weights and offers up to three modes: Exact, designed to reproduce the outputs of the unmodified model bit by bit; Fast, which sacrifices a small amount of numerical precision for speed; and Large, which reduces memory usage so that larger inputs can fit in GPU memory. The report states that Exact mode ran the forward pass of 14 structure prediction models an average of 1.6 times faster than the unmodified models on NVIDIA H100 GPUs, that Fast mode ran the 13 models it provides an average of 4.1 times faster, and that Large mode was an average of 3.4 times faster.
Modern structure prediction models, such as AlphaFold3, OpenFold3, and Boltz-2, consume a large portion of their runtime and memory for triangular attention and triangular multiplication, two operations whose cost is cubic in terms of the number of tokens. NVIDIA has developed dedicated cores for these operations, initially in cuEquivariance and more recently in the BioNeMo Inference Runtime. For the same bottleneck, Claude developed FlashPairformer v1, a set of custom GPU cores. The report indicates that FlashPairformer v1 performed triangular attention 2.7 times faster than industry standard cores at pair width 128, used by AlphaFold3 and Boltz-2, and 2.9 times faster at pair width 256, used by Protenix v2, while triangular multiplication was 1.7 times and 3.2 times faster, respectively, measured on a single NVIDIA H100.
Beyond the shared cores, Claude has made model-specific optimizations, including caching values that the original code recalculated, constant branch reduction with constant outputs, CUDA graph capture, and core fusion. The report states that the speedups in Rapid mode for structure prediction models ranged from 2.3x for OpenDDE to 6.4x for Chai-1, and that genomics and protein language models benefited from 1.2x to 5.0x in Exact mode.
In terms of accuracy, the report indicates that on FoldBench-Lite, a combined benchmark of 1,925 model-target pairs across 13 model configurations, the proportion of acceptably predicted interfaces was 54.8% in the default mode, compared to 55.0% for Exact, 54.5% for Rapid, and 54.2% for Mare, with no discernible combined change from zero. The limitations mentioned include measurements performed only on H100 GPUs, ColabFold speedups measured against version 1.6.1 before the project released its own fused cores, and Exact and Rapid modules using up to 3.2 times more memory than the default mode.
Large Module for Massive Biomolecular Systems
Anthropic said that Claude’s low-memory Large mode enables accurate modeling of biomolecular systems larger than 10,000 tokens (amino acids, nucleotides, and atoms from small molecules and ions) and inference on systems larger than 70,000 tokens, all on a single NVIDIA GPU node, which the paper describes as a previously unachievable task. Molecular machines accurately predicted in Large mode include human mitochondrial complex I, the chaperone complex TRiC, a proteasome, and a bacterial 70S ribosome, each closely matching the experimental structure with TM scores between 0.92 and 0.997, according to the report. Anthropic said that to its knowledge, these are among the largest structures ever accurately folded using structure prediction models; The report states that the AlphaFold3 work revealed an accurate prediction of the 40S ribosome of 7,663 tokens and that AlphaFold3 was trained on segments of at most 768 tokens.
To test the limits of the optimizations, Claude performed structure prediction on entire viral capsids and protein compartments ranging in size from about 31,000 to over 70,000 residues, using a node with eight B300 GPUs. The seven runs were completed, but each prediction collapsed into a tight ball, with TM scores between 0.08 and 0.14 where they were evaluated. The report assumes that the collapse reflects a generalization failure, as the ensembles are 40 to 90 times larger than AlphaFold3’s largest training set, and frames the result as an extension of what can be computed, rather than what the models have learned.
Binder design on a single GPU
In Anthropic’s previous work on binding design, Claude operated on a prompt of about 16,000 words, used sub-agents, and was able to spend up to $10,000 per target on the Modal AI infrastructure platform, equivalent to about 2,500 H100 GPU hours, over 24 hours. In the new configuration, a single Claude model was given an NVIDIA H200 GPU and 24 hours of real-time, operated on a prompt of about 1,100 words, and had no sub-agents or human guidance. Three models, Mythos 5.1, Mythos 5, and Opus 5, were each run on the 16 targets in the previous study.
The report states that median scores on ipSAE, an in silico score described as predictive of experimental binding, reached 0.785 for Opus 5, 0.781 for Mythos 5.1, and 0.739 for Mythos 5, compared to 0.749 for previous campaigns, with optimal design scores of 0.833, 0.825, and 0.813 versus 0.817. The single-GPU runs used about 100 times less GPU budget, and the post states that a combined spend of about $150 on GPUs and tokens achieved in silico performance equal to that of previous campaigns. The report notes that the designs resulting from these runs were not experimentally tested, and that the scores represent in silico measurements, not experimental results.
Open-source release conditions
Anthropic has published the code in a public GitHub repository containing 36 drop-in optimization kits, one for each core tool, with the original code released under the Apache 2.0 License; the initial public release was committed on September 17, 2026. The kits share a common off, accurate, fast, and large-scale vocabulary and are configured for NVIDIA H100 80 GB GPUs on Linux x86-64, with some kits also including configurations for A100, H200, B200, or B300. The repository specifies that the release is a reference, unmaintained, and non-contributory release, with pre-built Docker and Apptainer images planned within a week of the initial release.
Protein Design Competition with Adaptyv Bio
Anthropic said it, along with Adaptyv Bio, has selected five problems that are at the frontier of current protein design capabilities, including cross-species reactivity, pH sensitivity, and peptide-MHC specificity, as well as challenging targets such as GPCRs. According to the competition page , the two organizations are jointly sponsoring $1 million for experimental validation to test over 5,000 protein designs in Adaptyv’s automated lab at no cost to participants, and Anthropic is providing an additional $1 million in Claude credits. The Anthropic post states that Modal will provide up to $250,000 in computing credits and that Twist Bioscience will provide DNA for the competition.
Three participation tracks are offered: research labs and companies, with up to $50,000 in Claude credits for academic teams and up to $25,000 for industry teams, approximately 20 selected teams, and reserved lab testing for approximately 18 designs per challenge; individual researchers or teams of up to three people, with free Claude Max 20x access for the duration of the competition; and a self-supported open track where use of Claude is optional.
Applications are open from September 16 to September 24, 2026, and selected teams will be announced before September 28, 2026. Five weekly challenges run from September 28 to October 31, 2026, experimental validation is scheduled to be completed by November 30, 2026, and results will be published on Proteinbase on December 15, 2026. Participants retain ownership of their designs, all experimental results, including negative ones, will be published openly, and there is no cash prize; the page describes the schedule as planned and subject to change.
Comments