Contents
Anthropic spent four weeks having Claude rewrite the inference code for 36 open-source biomolecular modeling tools: AlphaFold-class structure predictors, protein design models, protein language models, and genomics tools.[1] The engineers supervising it had biology and modeling expertise but, by Anthropic’s own description, “no prior experience in inference optimization or kernel engineering.”[1] The resulting kernels beat Nvidia’s own optimized library on the single most expensive operation in these models by up to 3.2x, and a companion memory optimization let a research team fold a 10,000-plus-token protein complex on one GPU node, a job that used to require a cluster.[1] The same release also states, in one precise sentence, past what token count the underlying models simply stop working. That sentence is worth more than the speedup.
The operation that actually costs money
AlphaFold3-class models represent a protein or protein complex as a matrix of pairwise relationships between every residue, and update that matrix through a stack of operations called triangle attention and triangle multiplication, borrowed from AlphaFold2’s Evoformer. Both operations scale cubically with the number of tokens (residues, nucleotides, and atoms from bound small molecules or ions), because each update has to consider every pair of relationships in light of a third. Double the size of the system and the cost rises roughly eightfold. For any structure past a few hundred tokens, this is where the GPU time goes.[1]
FlashPairformer is Claude’s custom kernel implementation of these two operations. Anthropic reports it “outperforming the field standard on average by 2.7-2.9x on triangle attention and 1.7-3.2x on triangle multiplication,” where the field standard is Nvidia’s own cuEquivariance library, the thing teams already reach for to speed this up.[1] On top of the kernel work, Claude applied model-specific changes: caching computations that were being redundantly recomputed, and collapsing conditional branches that never actually varied at runtime into fixed constants.[1] None of this touches model weights or outputs in the modes meant to preserve them. It touches how the same math gets executed on the hardware.
Four correctness modes, not a promise
The part of this that’s actually a systems design lesson rather than a research result is how the code is packaged. Anthropic didn’t ship one faster version of each tool. The public repository gives each of the 36 optimized tools up to four modes: off runs the original, unmodified upstream code; exact is required to produce bit-identical outputs to off, just faster; fast allows small numeric differences within the tool’s own seed-to-seed variance, for more speed; and big trades some of that speed for the lowest peak GPU memory, to fit larger inputs.[2] Switching between them is a one-line change:
import os
os.environ['<KIT>_OPT'] = 'exact'
# run the tool's own script unchanged
The upstream tool’s code is kept in a stock/ directory that the optimization layer never edits, so the pinned original is always there to diff against.[2] This is what makes it reasonable to trust performance-critical code an LLM wrote without re-deriving the kernels yourself: exact mode isn’t a claim, it’s a runnable, falsifiable check, and every prebuilt binary in the repo is SHA-256 digest-verified before it loads.[2] If your own team is going to let an agent touch inference code that other people’s results depend on, this is closer to the right shape than a test suite you have to trust was thorough: a mode that structurally cannot change the answer, sitting next to the one that can.
What “big” mode actually bought, and where it stopped
The memory optimization is where the honesty gets specific. Claude built a low-memory big mode that made “the accurate modeling of systems larger than 10,000 tokens” possible, and got as far as “successful inference on systems larger than 70,000 tokens using just one NVIDIA GPU node.”[1] Using this, Anthropic folded human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial ribosome, each closely matching its experimentally determined structure, with complex I and the ribosome each exceeding 10,000 tokens.[1] For comparison, AlphaFold3’s own previous largest accurate prediction, the 40S ribosome, was 7,663 tokens.[1] That’s a real jump in what fits in a single node’s memory, and it stayed inside the models’ learned distribution.
Then Anthropic pushed on purpose. They asked Claude to predict viral capsids and protein compartments ranging from over 31,000 to over 70,000 tokens, deliberately larger than anything attempted before.[1] The result: “these systems are not predicted correctly,” and “predicted structures collapse, suggesting a lack of generalization nearly two orders of magnitude beyond the training context.”[1] The two ranges overlap, 31,000 to 70,000 tokens versus “larger than 70,000,” which means the failure isn’t purely a token-count cliff. The ribosome and complex I succeeded because they’re the kind of assembly, real cellular machinery, that resembles what these models were trained on. The capsids and compartments failed at similar or smaller scale because they’re structurally unlike anything in that training distribution. Big mode solved the memory problem. It did nothing for, and couldn’t have done anything for, the generalization problem, because those are different bottlenecks that happen to both show up as “can we run this.”
The economics, and what changes because of them
The same optimization work also cut the cost of de novo protein binder design by close to two orders of magnitude. Anthropic’s earlier approach gave Claude an approximately 16,000-word prompt, sub-agents, and up to $10,000 per target on Modal, roughly 2,500 Nvidia H100 GPU hours.[1] The new setup gives a single Claude model one Nvidia H200, 24 hours of wall time, and a roughly 1,100-word prompt with a reference sheet for the pre-installed tools. Averaged over 16 targets, the median and highest-scoring designs from three Claude models, Mythos 5.1, Mythos 5, and Opus 5, matched the binding-confidence scores (ipSAE) of the earlier, far more expensive campaigns, for a combined spend of about $150 in GPU and token costs.[1]
That gap, $150 versus $10,000 for comparable in silico results, is what makes the generalization boundary matter practically rather than academically. Cheap iteration means more people will run these tools past the sizes anyone has validated, not fewer, because the cost of trying no longer discourages it. The corresponding decision for anyone building on this work is not to distrust the speedup, which is verifiable through exact mode and now open-sourced at github.com/anthropics/uplifting-biomolecular-modeling.[2] It’s to treat “runs successfully on one GPU node” and “predicted correctly” as two separate claims that this exact release shows can diverge, and to gate scale the way Anthropic did here: test past the known boundary on purpose, on structures you can independently verify, before trusting an answer you can’t.
Sources
[1] https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling: “How Claude is uplifting biomolecular modeling,” Anthropic, September 22, 2026
[2] https://github.com/anthropics/uplifting-biomolecular-modeling: Anthropic, optimization kit repository and README, September 2026