Anthropic says an internal research model, working without a specialized coding setup, rewrote more than 30 open-source biology models in under four weeks and made them run roughly four times faster on average. The company disclosed the work in a research post published Thursday. Alongside it, Anthropic announced a new protein design competition with Adaptyv Bio, committing as much as $1 million in Claude credits to the prize pool.
The models in question predict protein structures and design new molecules, tools that labs such as those behind AlphaFold3 and Boltz-2 already use for drug discovery and basic biology. Anthropic had two of its own staff, both experienced in biomolecular modeling but new to kernel engineering, supervise the project rather than hand it to specialists in low-level GPU optimization.
The centerpiece is a set of custom kernels the company calls FlashPairformer. Triangle attention and triangle multiplication, the calculations that eat most of the runtime and memory budget in these models, were its target. Anthropic reports the kernels beat NVIDIA’s existing cuEquivariance library, the current field standard, by 2.7 to 2.9 times on one of those calculations and 1.7 to 3.2 times on the other, depending on model configuration. These are Anthropic’s own benchmark numbers. The post does not cite independent testing of FlashPairformer.
A second result addresses memory rather than speed. Anthropic says a new low-memory “Big” mode lets these models accurately predict biomolecular assemblies larger than 10,000 amino acids, nucleotides, and atoms, and run inference (with degraded accuracy) on structures beyond 70,000, all on a single NVIDIA GPU node. Previously, systems that size needed compute spread across multiple nodes. The company folded structures including a bacterial ribosome and a human mitochondrial complex using the new mode and reports each closely matched its known experimental structure.
The protein design result is the sharpest comparison to Anthropic’s earlier work. In an April project, the company gave Claude a budget of roughly $10,000 for each target, equivalent to about 2,500 NVIDIA H100 GPU hours, alongside a 16,000-word prompt and sub-agents to design binding proteins. The new test cut all of that down: a single Claude model, one NVIDIA H200 GPU, a 1,100-word prompt, and no sub-agents produced designs that scored about the same on an in silico binding metric called ipSAE, for roughly $150 in combined GPU and token spend. That is a claimed reduction of about two orders of magnitude in compute for comparable results on paper, not yet in a wet lab.
ipSAE is a computational proxy for binding, not a measurement of whether a protein actually works outside a simulation. Anthropic’s plan to close that gap is the protein design competition it is co-sponsoring with Adaptyv Bio, the company that has run open protein design contests before. Five problems, including species cross-reactivity and difficult GPCR targets, will be scored against more than 5,000 submitted designs with actual wet lab validation, funded by the $1 million in Claude credits plus $250,000 in Modal compute credits and DNA synthesis from Twist Bioscience.
Anthropic is also open-sourcing the optimized code and a technical report, and it opened its Life Sciences Verification Program to public beta the same day, after enrolling an initial group of research organizations.
For biotech teams running structure prediction or protein design pipelines, the FlashPairformer kernels and the low-memory mode are worth testing directly, since both are open source now rather than locked inside Anthropic’s infrastructure. The wet lab results from the Adaptyv competition, expected only after the entry period closes, will be the first outside check on whether the compute savings hold up past a simulation.
Based on Anthropic’s research post “How Claude is uplifting biomolecular modeling,” published September 18, 2026.
