Anthropic opens source code: 36 optimized packages for biomolecular modeling

Anthropic opens source code: 36 optimized packages for biomolecular modeling

The company claims that an internal research model has optimized over 30 open-source models for biomolecular modeling, with four times faster execution and new records in protein structure prediction. Now they are inviting the rest of the field to prove it.

Anthropic published research results, a technical report, and open-source code on September 17, 2026, for an extensive optimization effort in which an internal, general-purpose research model — that is, Claude — has optimized more than 30 open-source models for biological tasks. In addition, the company launched a protein design competition together with Adaptyv Bio.

What was actually done

According to reporting from Unite.AI, which builds on Anthropic's technical report, the model produced optimized packages for 36 model implementations covering over 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).

What distinguishes this from previous "AI helps scientists" stories is who provided supervision. Anthropic states that Claude was primarily supervised by two employees on the technical staff — with experience in biomolecular modeling but no experience in inference optimization or kernel engineering. That is a deliberate point: This was not an elite GPU team getting assistance, but domain experts without kernel expertise who had an AI model do the heavy lifting.

The numbers — with the caveat that they come from Anthropic itself

The company claims that Claude made the tasks run about four times faster on average with minimal loss in precision, and almost twice as fast with identical results.

The most striking result is FlashPairformer, a set of custom GPU kernels for triangle attention and triangle multiplication, developed in collaboration with Claude for Anthropic's own work on optimizing inference for structure prediction models. The company claims this sets a new standard, beating the field's established standard kernels by 2.7–2.9 times on triangle attention and 1.7–3.2 times on triangle multiplication, depending on model configuration. These operations are among the computationally heaviest in modern protein structure models, so improvements here are directly relevant to the field.

It is worth emphasizing: Almost all performance numbers here come from Anthropic's own report, and no independent third-party benchmarking has been provided. The code has, however, been made public, so the field can now verify the claims themselves.

"Big mode": large complexes on one GPU node

The memory-efficient "Big mode" configuration is said, according to the report, to enable accurate modeling of biomolecular systems of over 10,000 tokens (amino acids, nucleotides, and atoms from small molecules and ions) and inference on systems over 70,000 tokens — all on a single NVIDIA GPU node, which Anthropic's post describes as a task previously out of reach for the field.

The report cites accurate predictions of molecular machines such as human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial 70S ribosome, all of which are said to closely match experimental structures with TM-scores from 0.92 to 0.997.

Competition and cost reduction

Previously, Anthropic's binder design pipeline cost up to $10,000 per target (around 2,500 H100 GPU hours). With the new optimizations and simplifications in the agent-based protein design approach, the company claims that comparable in silico performance is now achieved with two orders of magnitude fewer GPU hours.

To strengthen the community, Anthropic is co-sponsoring a protein design competition together with Adaptyv Bio. The competition consists of five challenging problems at the frontier of current capabilities. With contributions from Modal and Twist Bioscience, the companies are putting up up to $1 million in Claude credits, $250,000 in Modal compute credits, as well as wet lab validation for over 5,000 designs.

Anthropic also points to an earlier result where Claude Mythos 5.1 accelerated seven open-source biology models, making them run up to 2.5 times faster.

The limitations Anthropic itself points out

The report is surprisingly open about what did not work. On the pooled FoldBench-Lite benchmark — 1,925 model–target pairs across 13 model configurations — the proportion of acceptably predicted interfaces was 54.8 percent with the standard setup, versus 55.0 for Exact, 54.5 for Fast, and 54.2 for Big. No change was distinguishable from zero. In other words: The optimizations made things faster, not better.

Furthermore, all seven extrapolation tests on viral capsids and protein chambers (around 31,000 to over 70,000 residues) collapsed into compact balls, with TM-scores of 0.08–0.14 where measurable. The report assumes that the collapse reflects an inability to generalize, since the assemblies are 40 to 90 times larger than AlphaFold3's largest training crop.

It is also worth noting that the measurements were only performed on H100 GPUs, and that the Exact and Fast modes use more memory.

What this shows — and what it does not

Anthropic itself frames the result as an extension of what can be computed, not of what the models have learned. Big mode brings large complexes within reach of computation, but the capsid tests show that the models still cannot predict structures far outside their training data.

The unresolved question is how much of this holds up outside Anthropic's own measurements. Now that the code is public, it is up to the community — and the competition with Adaptyv Bio — to find out.

Sources: Anthropic, Unite.AI