Anthropic Opens Claude-Built Biology Tool Optimizations It Says Run 4x Faster

Anthropic has released open-source code for optimizations it says Claude developed across more than 30 biomolecular modeling tools. The company reports roughly fourfold average speed gains with minimal precision loss, a result that could lower the computing barrier for scientists using software to predict and design molecules.

The work centers on programs for tasks including protein structure prediction, protein design, genomics and protein-language modeling. Anthropic says Claude completed the optimizations in just under four weeks under supervision from two technical staff members experienced in biomolecular modeling.

Speeding up the costly geometry

Modern structure-prediction systems spend much of their time on triangle attention and triangle multiplication, calculations that help model molecular geometry. Their cost grows cubically: doubling a system’s size can require eight times the time and memory. Anthropic says its new FlashPairformer kernels accelerate those two operations by 2.7 to 2.9 times and 1.7 to 3.2 times, respectively, depending on the model configuration.

The shared kernels are only part of the release. Anthropic says it also used model-specific changes, such as caching redundant calculations and simplifying inactive code paths. Across the optimized models, it reports roughly 4x speed-ups with minimal precision loss and nearly 2x speed-ups when outputs remain identical.

A memory-saving mode, with a clear boundary

The release also includes a low-memory setting called Big mode. Anthropic says it accurately modeled systems above 10,000 tokens on one NVIDIA GPU node, including mitochondrial complex I, a bacterial ribosome, a proteasome and the TRiC chaperone complex. That matters because large molecular machines can be difficult to model without spreading inference across multiple machines.

But successfully running a larger input is not the same as predicting it correctly. Anthropic says it processed viral capsids and protein compartments ranging from more than 31,000 to more than 70,000 tokens on a single eight-GPU B300 node. The predicted structures collapsed, which Anthropic says suggests the models did not generalize at that scale.

Lower-cost design is still an in-silico result

Anthropic also tested the optimized tools in a de novo protein-binder workflow. It says a single Claude model running for 24 hours on one NVIDIA H200 reached binding scores comparable to its earlier campaigns across 16 targets, while using about two orders of magnitude fewer GPU hours. Those are computational scores, not evidence that the designed binders work in a laboratory.

That distinction defines the practical value of this release. Faster software can let researchers explore more candidate molecules or larger systems with a fixed compute budget. It cannot, on its own, remove the need to determine whether a prediction is accurate at the scales and tasks where it will be used.