Claude Sped Up More Than 30 Open-Source Biomolecular Models by Roughly 4x in Under Four Weeks

Anthropic published results on September 17, 2026 showing that Claude optimized more than 30 open-source biomolecular models in just under four weeks, speeding them up roughly 4x on average. The 4x figure holds when a minimal amount of precision is sacrificed, and the speedup is nearly 2x when outputs are identical, while a low-memory "Big" mode Claude created enables accurate prediction of biomolecular systems larger than 10,000 tokens on a single NVIDIA GPU node. The same post reports de novo binder design against 16 targets using one NVIDIA H200 and 24 hours of wall time at roughly two orders of magnitude fewer GPU hours, plus a protein design competition co-sponsored with Adaptyv Bio. ASAP separates the conditions that have to be read before the speedup multiples.

Two triangle operations consume most of the structure prediction budget

AlphaFold3, OpenFold3 and Boltz-2 spend much of their runtime and memory on triangle attention and triangle multiplication, and those two operations are cubic, so doubling the size of a system costs 8x more and tripling it costs 27x more. Both operations act on triplets of tokens, and they make it possible to model the geometry of biomolecular systems at that cubic price in time and memory. Tokens here means amino acids, nucleotides, and atoms from small molecules and ions.

That cost structure already made these two operations the target of dedicated kernel work. NVIDIA's cuEquivariance came first, and more recently NVIDIA's BioNeMo Inference Runtime. Kernels are low-level software translation layers for accelerated computing hardware such as GPUs, which means rewriting the expensive operations against the specific characteristics of the chip.

What Anthropic built with Claude belongs to that layer. The target is the core of the Pairformer architecture underlying state-of-the-art structure prediction models, and the comparison baseline is a field standard that engineers have been hand-tuning for years.

FlashPairformer takes NVIDIA's kernels as its baseline

FlashPairformer, the set of custom kernels Anthropic developed with Claude, outperforms the field standard on average by 2.7-2.9x on triangle attention and 1.7-3.2x on triangle multiplication, with the range depending on model configuration. Anthropic records the result as a new state of the art. The comparison target is not a competing academic kernel but NVIDIA's own dedicated work, which is what gives these numbers their weight.

Model-specific optimizations ran alongside the transferable kernels. Anthropic cites changes such as caching redundant recomputed work and simplifying dead branches into their constant outputs. Together the two tracks accelerated the structure prediction models by 4x on average, and Anthropic states that for each model it confirmed the accelerated versions did not impact performance on the downstream task such as structure prediction.

This is where the conditions have to be kept separate when quoting. The roughly 4x average across all 30-plus models holds while sacrificing a minimal amount of precision, and the figure is nearly 2x with identical outputs. Narrowed to structure prediction models the pair becomes roughly 4x and roughly 1.6x. Anthropic also notes that ColabFold 1.6.3's concurrently-released optional fast kernels are not yet benchmarked here.

Big mode unlocked size rather than speed

The low-memory Big mode Claude created enables accurate modeling of systems larger than 10,000 tokens on a single NVIDIA GPU node, and Anthropic names 4 molecular machines folded with it. They are human mitochondrial complex I, the TRiC chaperone complex, a proteasome and a bacterial ribosome, each closely matching its experimentally determined structure. Anthropic contrasts complex I and the 70S ribosome, at more than 10,000 tokens each, with the 40S ribosome predicted accurately by AlphaFold3 at 7,663 tokens.

The accuracy threshold is stated alongside. A predicted interface counts as acceptable when its DockQ score is greater than 0.23. Anthropic reports that the fast modes developed for structure prediction are statistically indistinguishable from the default settings across a pooled set of biomolecular interfaces.

Much of the work in a cell is done by systems of exactly this size. The ribosome that builds proteins, the respiratory complexes that power the cell, and the chaperones that help other proteins fold are each built from dozens of components whose function depends on how they fit together. Predicting structures at that scale has typically required inference spread across multiple GPU nodes, which put it out of reach for most molecular biologists.

Four weeks and two supervisors matter more than the 4x

Anthropic states that 2 members of its technical staff with no prior experience in inference optimization or kernel engineering supervised Claude through the acceleration of more than 30 models in just under four weeks. Both are experienced in biomolecular modeling. As Anthropic itself notes, this kind of optimization normally takes an experienced team of engineers weeks per model, and the work often does not transfer between models.

Judged on the multiples alone the result is impressive without being unfamiliar. Kernel optimization is a field where many teams have delivered, and 2.7x or 4x sits inside the range humans reach given enough time. What actually changed is the staffing required to get there. Two researchers with domain knowledge but no low-level performance engineering background, acting only as supervisors, covering 30 models in four weeks is the part that is harder to reproduce and therefore the more important claim.

In practice that distinction moves the unit of the decision. Kernel work has been a line item requiring a dedicated hire or an outside contract, and the recurring question was whether a single model justified that cost. If a two-supervisor arrangement holds, the unit shifts from one model to the entire set of models a team runs. What Anthropic published is one case obtained with an internal general-purpose research model, and evidence that the same arrangement repeats across other codebases and other domains is not part of this post.

The collapsed capsid predictions draw the boundary line

Claude generated predictions of entire viral capsids and protein compartments ranging from more than 31,000 to more than 70,000 tokens on a single NVIDIA B300 node with 8 GPUs, and Anthropic states that these structures are not predicted correctly and collapse. Those systems sit nearly two orders of magnitude beyond the training context of the models, and Anthropic describes the outcome as a lack of generalization. The capability runs used a single trunk pass with no recycles as proof of concept.

The value of this section is that the failure was not removed from the post. Big mode delivered two different things, and quoting only one of them makes the claim false. Around 10,000 tokens the predictions are accurate, and around 70,000 tokens memory holds but the result falls apart. What was solved is the compute threshold, and what remains unsolved is the generalization limit of the models themselves.

That split identifies where the next bottleneck sits. With weights and training data unchanged, lifting the memory constraint moved the point where the curve breaks from hardware to training context. Read together, Anthropic's note that the models generalize nearly 1.5 orders of magnitude beyond their training context and its note that they collapse nearly two orders of magnitude beyond it bracket roughly where the boundary lies. For a team that wants to work on large molecular machines, the practical next question is not whether to add nodes but when structure prediction models trained on longer contexts arrive.

Cutting GPU hours by 100x changes who can participate

Three Claude models designed de novo binders against 16 targets with one NVIDIA H200 and 24 hours of wall time, reaching in silico performance matching the earlier campaigns at a combined spend of approximately $150 on GPUs and tokens. In the earlier work Anthropic gave Claude an approximately 16,000-word prompt that encouraged sub-agents and allowed up to $10,000 per target on Modal, roughly 2,500 NVIDIA H100 GPU hours, within a 24-hour span. This time a single Claude model received a prompt of about 1,100 words and a reference sheet for pre-installed tools, with no sub-agents and no human steering the designs.

The scoring metric is ipSAE, an in silico score that has been shown to be predictive of binding in the wet lab. Averaged over 16 targets, the median-scoring and highest-scoring designs from all three Claude models evaluated achieve approximately the same ipSAE values as the earlier Mythos 5.1 campaigns. The three models are Mythos 5.1, Mythos 5 and Opus 5.

The word to handle carefully is in silico. What was matched is the computational score of the earlier campaigns, and the post contains no statement that wet lab binding rates were measured again for this run. Even with ipSAE described as predictive of binding, a tie on the metric and a tie at the bench are two different claims.

With that caveat attached, the implication still stands. When work that cost $10,000 per target drops to roughly $150, the size of the population that can do it changes. As Anthropic acknowledged in the earlier post, the previous approach took more resources than would be available to the vast majority of protein designers. One GPU and one day brings the work inside the range of a university lab or a small biotech running its own targets. Whether the open-sourced optimization code actually lowers that threshold is something outside reproduction has to confirm.

The competition and verification program push the measurement outside

Anthropic is committing up to $1 million in Claude credits and $250,000 in Modal compute credits, plus wet lab validation for over 5,000 designs, to a protein design competition run with Adaptyv Bio. Twist Bioscience is providing DNA. The two teams jointly selected five problems, including species cross-reactivity, pH-sensitivity and peptide-MHC specificity, along with difficult targets such as GPCRs.

That structure connects back to the caveat above. A claim of parity made on an in silico score now has a procedure attached in which more than 5,000 community designs get synthesized and measured. Separately, Anthropic states that it enrolled the first group of organizations in its Life Sciences Verification Program and opened the program in public beta the same day.

For research groups outside the competition, the immediately usable part of this announcement is the code rather than the contest. The optimized model code is open-sourced and a technical report accompanies it, so attaching the same kernels to an existing pipeline and measuring the speedup independently can start today. Re-deriving the numbers on your own hardware and your own targets is safer than transplanting the published 4x average into your environment as an expectation.

Source: How Claude is uplifting biomolecular modeling (Anthropic, September 17, 2026)