Anthropic Reports Claude Optimized 30+ Open-Source Biomolecular Models

Anthropic published a research post on September 17, 2026, reporting that Claude optimized more than 30 open-source biomolecular models in just under four weeks, which the company said sped the models up roughly 4x on average. Anthropic open-sourced the optimized code and announced a protein design competition co-sponsored with Adaptyv Bio.

36 Optimized Packages and FlashPairformer Kernels

According to the research post and an accompanying technical report, an internal general-purpose research model produced optimized packages for 36 model implementations covering more than 30 open-source models across six families: co-folding and structure prediction (14 packages), hallucination (3), structure generation (6), inverse folding (3), genomics (7), and protein language models (3). Anthropic said Claude was supervised primarily by two members of its technical staff with experience in biomolecular modeling but none 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 unmodified model’s outputs bit for bit; Fast, which trades a small amount of numerical precision for speed; and Big, which reduces memory use so larger inputs fit in GPU memory. The report states that Exact mode ran the forward pass of 14 structure prediction models on average 1.6x faster than the unmodified models on NVIDIA H100 GPUs, that Fast mode ran the 13 models offering it on average 4.1x faster, and that Big mode averaged 3.4x faster.

Modern structure prediction models such as AlphaFold3, OpenFold3, and Boltz-2 spend much of their runtime and memory on triangle attention and triangle multiplication, two operations whose cost is cubic in the number of tokens. NVIDIA has developed dedicated kernels for these operations, first in cuEquivariance and more recently in the BioNeMo Inference Runtime. For the same bottleneck, Claude developed FlashPairformer v1, a set of custom GPU kernels. The report states that FlashPairformer v1 ran triangle attention 2.7x faster than the field-standard kernels at pair width 128, used by AlphaFold3 and Boltz-2, and 2.9x faster at pair width 256, used by Protenix v2, with triangle multiplication 1.7x and 3.2x faster respectively, measured on a single NVIDIA H100.

Beyond the shared kernels, Claude produced optimizations specific to each model, including memoizing intermediates the original code recomputed, constant-folding branches with constant outputs, CUDA graph capture, and kernel fusion. The report states that Fast-mode speed-ups for the structure prediction models ranged from 2.3x for OpenDDE to 6.4x for Chai-1, and that genomics and protein language models gained 1.2x to 5.0x in Exact mode.

On accuracy, the report states that across FoldBench-Lite, a pooled benchmark of 1,925 model–target pairs over 13 model configurations, the share of acceptably predicted interfaces was 54.8% at default versus 55.0% for Exact, 54.5% for Fast, and 54.2% for Big, with no pooled change distinguishable from zero. Stated limitations include measurements taken only on H100 GPUs, ColabFold speed-ups measured against version 1.6.1 before that project released its own fused kernels, and Exact and Fast modes using up to 3.2x more memory than default.

Big Mode for Massive Biomolecular Systems

Anthropic said Claude’s low-memory Big 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 post describes as a previously out-of-reach task. Molecular machines predicted accurately in Big mode include human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial 70S ribosome, each closely matching its experimental structure with TM-scores from 0.92 to 0.997, according to the report. Anthropic said these are, to its knowledge, among the largest structures ever folded accurately using structure prediction models; the report notes that the AlphaFold3 paper highlighted an accurate 40S ribosome prediction of 7,663 tokens and that AlphaFold3 was trained on crops of at most 768 tokens.

To test the limits of the optimizations, Claude ran structure prediction on entire viral capsids and protein compartments ranging from roughly 31,000 to more than 70,000 residues using one node of eight B300 GPUs. All seven runs completed, but every prediction collapsed into a compact ball, with TM-scores of 0.08 to 0.14 where scored. The report hypothesizes that the collapse reflects a failure to generalize, because the assemblies are 40 to 90 times larger than AlphaFold3’s largest training crop, and frames the result as extending what can be computed rather than what the models have learned.

Binder Design on a Single GPU

In Anthropic’s earlier binder-design work, Claude operated from a roughly 16,000-word prompt, used sub-agents, and could spend up to $10,000 per target on the AI infrastructure platform Modal, roughly 2,500 H100 GPU hours, within 24 hours. In the new setup, a single Claude model received one NVIDIA H200 GPU and 24 hours of wall time, worked from a prompt of about 1,100 words, and had no sub-agents and no human steering. Three models, Mythos 5.1, Mythos 5, and Opus 5, were each run against the 16 targets of the earlier study.

The report states that median scores on ipSAE, an in silico score it describes as predictive of experimental binding, reached 0.785 for Opus 5, 0.781 for Mythos 5.1, and 0.739 for Mythos 5, against 0.749 for the earlier campaigns, with best-design scores of 0.833, 0.825, and 0.813 versus 0.817. The single-GPU runs used a GPU budget about 100 times smaller, and the post states that a combined spend of approximately $150 on GPUs and tokens achieved in silico performance matching the earlier campaigns. The report states that designs from these runs have not been tested experimentally and that the scores are in silico measurements, not experimental results.

Open-Source Release Terms

Anthropic released the code in a public GitHub repository containing 36 drop-in optimization kits, one per upstream tool, with the original code under Apache License 2.0; the initial public release was committed on September 17, 2026. The kits share a common off, exact, fast, and big mode vocabulary and are configured for NVIDIA H100 80 GB GPUs on Linux x86-64, with some kits also shipping A100, H200, B200, or B300 configurations. The repository states that the release is a reference release that is not maintained and not accepting contributions, and that prebuilt Docker and Apptainer images are planned within a week of the initial release.

Protein Design Competition With Adaptyv Bio

Anthropic said it and Adaptyv Bio selected five problems at the frontier of current protein design capabilities, including species cross-reactivity, pH-sensitivity, and peptide-MHC specificity, as well as difficult targets such as GPCRs. According to the competition page, the two organizations are jointly sponsoring $1 million in experimental validation to test more than 5,000 protein designs in Adaptyv’s automated lab at no cost to participants, with Anthropic providing an additional $1 million in Claude credits. Anthropic’s post states that Modal will provide up to $250,000 in compute 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, roughly 20 teams selected, and reserved lab testing of up to approximately 18 designs per challenge; individual researchers or teams of up to three, with complimentary Claude Max 20x access during the competition; and a self-supported open track in which Claude use is optional.

Applications run from September 16 to September 24, 2026, and selected teams are to be notified before September 28, 2026. Five weekly challenges run from September 28 to October 31, 2026, experimental validation is scheduled to complete by November 30, 2026, and results are to be published on Proteinbase on December 15, 2026. Participants retain ownership of their designs, all experimental results including negative results will be published openly, and there is no cash prize; the page describes the schedule as planned and subject to change.