Anthropic says Claude made biology models roughly four times faster

Anthropic says Claude has optimised more than 30 open-source biology models in under four weeks, with average speed-ups of roughly four times and a new low-memory mode for much larger molecular systems. The practical significance is clear: some protein-structure and design work could require far less computing, although the results remain Anthropic’s own research claims rather than independent validation.

Anthropic Watch analysis

What happened

In a research post published on 17 September, Anthropic says Claude helped develop FlashPairformer, custom kernels for expensive operations used by biomolecular structure-prediction models. The company reports average speed-ups of about four times, with nearly twice the speed in tests where outputs were identical. Anthropic says it checked that the faster versions did not harm downstream tasks such as structure prediction.

The work covered more than 30 open-source models across structure prediction, protein design, protein-language modelling and genomics. Anthropic says it is releasing the optimised code, giving researchers something more useful than a dazzling benchmark slide: a chance to inspect and run the work themselves.

The post also describes a low-memory “Big” mode that can run models for biomolecular systems larger than 10,000 tokens on one NVIDIA GPU node. Anthropic says successful examples included a mitochondrial complex, a chaperone complex, a proteasome and a bacterial ribosome. It reports that even larger tests, above 31,000 tokens, remained inaccurate, so the frontier has not suddenly acquired manners.

Why it matters

Biology models are often limited not only by model quality but by memory, runtime and the price of the hardware needed to use them. Lowering those costs could make structure prediction and protein design more accessible to researchers who cannot justify multi-GPU infrastructure.

Anthropic also says its earlier protein-design work used up to $10,000 per target and about 2,500 NVIDIA H100 GPU hours. In the newer experiment, it reports comparable in-silico performance with roughly two orders of magnitude fewer GPU hours. That is an important claim, but it is not the same as proving a designed protein will work in a laboratory.

Our read

This is a substantial AI-for-science story because the reported improvement is in the plumbing that determines who can afford to do the science. The open-source release matters more than the headline speed-up, because other researchers can test whether the gains travel beyond Anthropic’s chosen models and workloads.

Treat the results as promising engineering evidence, not a shortcut to new medicines. Wet-lab validation, reproducible benchmarks and performance on unfamiliar systems will decide whether this is a durable change or an unusually polished internal demo.

What to watch

• Whether independent researchers reproduce the reported speed-ups on the released code.
• How much precision changes across different models and biological tasks.
• Whether the low-memory mode works reliably on ordinary research hardware.
• Results from the protein-design competition and its wet-lab validation. Anthropic’s report is available here.

Discussion spark: If Anthropic’s optimisations hold up independently, should the field prioritise cheaper and more reproducible biology models over chasing larger ones?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.