Claude Optimized 30+ Biomolecular Models: 4x Average Speedups and a Single-Node Big Mode
Anthropic has open-sourced Claude-assisted inference optimizations for more than 30 biomolecular deep-learning models after a project lasting just under four weeks. The company reports roughly 4x average acceleration with minimal precision loss across the optimized workloads and nearly 2x acceleration when requiring identical outputs. The work spans structure prediction, protein design, protein language models and genomics.
The September 17 release also includes FlashPairformer, custom GPU kernels for the triangle-attention and triangle-multiplication operations used by modern biomolecular structure models. Anthropic reports 2.7–2.9x acceleration for triangle attention and 1.7–3.2x for triangle multiplication against the field-standard implementation used in its comparison, depending on model configuration.
A separate low-memory Big mode extends the size of systems that can be processed on one GPU node. Anthropic reports accurate predictions for biomolecular systems above 10,000 tokens and successful inference runs above 70,000 tokens on a single eight-GPU NVIDIA B300 node. Accuracy and executable scale diverge at the extreme end: the 31,000–70,000-token proof-of-concept structures ran successfully, while their predicted structures collapsed and were inaccurate.
What Claude changed
The project optimized inference software around existing specialist biology models. Anthropic says Claude developed transferable GPU kernels and model-specific optimizations including caching redundant computation and simplifying branches whose outputs were effectively constant.
The engineering target matters because Pairformer-style structure models spend substantial compute and memory on triangle operations over token triplets. Anthropic describes these operations as cubic in runtime and memory: doubling system size can require roughly eight times the resources for the affected operation, while tripling it can require roughly 27 times.
Two Anthropic technical staff members supervised the work. They had biomolecular-modeling experience and, according to Anthropic, entered the project without prior inference-optimization or kernel-engineering experience. The company says Claude completed optimizations across more than 30 open-source models in under four weeks.
Anthropic has published the optimized implementations so researchers can inspect and run the code and directly test the benchmark claims.
The benchmark results
Across more than a dozen structure-prediction models, Anthropic reports an average speedup of roughly 4x in its fast configurations with minimal precision reduction. The identical-output configuration is more conservative; Anthropic reports roughly 1.6x across the structure-model subset shown in its benchmark and describes the broader optimized set as approaching 2x when outputs are held identical.
The company evaluated downstream structure quality alongside kernel timing. For the fast structure-prediction modes, it reports results statistically indistinguishable from default settings across a pooled set of biomolecular interfaces, using DockQ greater than 0.23 as the acceptable-interface threshold.
The published benchmark is vendor research from Anthropic. The released code makes the performance claims testable on other hardware and datasets, which is particularly useful because kernel speedups can change with GPU generation, tensor shapes, model configuration and batching.
Big mode trades memory pressure for much larger inference jobs
Anthropic's Big mode focuses on memory use for large molecular machines such as ribosomes, respiratory complexes and chaperones. It reports accurate predictions above 10,000 tokens on one NVIDIA GPU node for examples including human mitochondrial complex I, the TRiC chaperone complex, a proteasome and a bacterial ribosome.
The company then pushed the optimized stack further on one eight-GPU B300 node. Inference completed for viral capsids and protein compartments ranging from more than 31,000 to more than 70,000 tokens. Those extreme-scale structures were inaccurate, so the result establishes an execution-capacity milestone without validating structure-prediction quality at 70,000 tokens.
That distinction is operationally important for scientific-computing teams. Lower memory requirements can make previously multi-node experiments executable on one node. Usable scientific output still depends on the model generalizing at the requested biological scale.
Protein-design compute fell by roughly two orders of magnitude
Anthropic also reran its agentic de novo protein-binder workflow with the optimized models. Its earlier campaign allowed Claude to use sub-agents and spend up to $10,000 per target, equivalent to roughly 2,500 H100 GPU-hours on Modal.
In the new setup, one Claude model received one NVIDIA H200, 24 hours of wall time, a shorter tool-reference prompt and no sub-agents. Anthropic tested Mythos 5.1, Mythos 5 and Opus 5 across 16 targets. It reports approximately the same in-silico ipSAE binding scores as the earlier Mythos 5.1 campaigns while using about 100 times fewer GPU-hours.
Anthropic estimates that approximately $150 combined GPU and Claude-token spend can match the in-silico performance level of its earlier campaigns. The comparison is an in-silico result; wet-lab binding performance is a separate validation step.
Why the release matters for scientific AI infrastructure
The immediate value is software efficiency. Biomolecular research increasingly combines large general-purpose AI agents with specialist GPU models for structure prediction, sequence modeling and molecular design. Faster kernels and lower memory consumption can reduce the number of accelerator-hours required per experiment and increase the size of systems that fit within a fixed node.
The project also demonstrates a concrete use of frontier coding models in performance engineering: generating and integrating specialized GPU kernels across heterogeneous scientific codebases under expert supervision. The released implementations provide a practical basis for measuring how those gains transfer beyond Anthropic's benchmark environment.
For infrastructure planning, the strongest result is the combination of roughly 4x average fast-mode acceleration, an identical-output path around 1.6–2x depending on the evaluated subset, and successful single-node execution at much larger token counts. Teams evaluating the code should benchmark their own models, GPU generation and accuracy thresholds before translating those ratios directly into capacity or cost forecasts.
Open-source code and competition
Anthropic released the optimized code alongside a protein-design competition with Adaptyv Bio. The program includes up to $1 million in Claude credits, up to $250,000 in Modal compute credits, and wet-lab validation for more than 5,000 designs across five protein-design challenges.
The open-source release gives developers and research-computing teams inspectable kernels and model-specific optimization packages that can be reproduced and profiled against existing biomolecular pipelines. Independent reproduction will determine how broadly the reported gains carry across hardware, model versions and real laboratory workflows.
Sources
- Anthropic research — How Claude is uplifting biomolecular modeling: https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling
- Anthropic open-source optimization implementations: https://github.com/anthropics/uplifting-biomolecular-modeling