Claude speeds up 30+ open-source biology models about 4x and folds 10,000+ token complexes on one GPU node
Confirmed
The takeaway
On Sept 17, 2026 Anthropic reported that Claude, working in Claude Science, optimized more than 30 open-source biomolecular models (structure prediction, protein design, genomics, protein language models) in under four weeks.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 30 September 2026
Your AI and this story
- GPT-6.2 Astra109 days after its cutoff
- Claude Opus 5.579 days after its cutoff
- Gemini 4 Hydrogen170 days after its cutoff
- Grok 4.7109 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 79 days before it.
Key facts
- 30+ models optimized in just under four weeks by an internal general-purpose research model, supervised by two staff with no prior kernel-engineering experience
- Average speedup ~4x with minimal loss of precision, ~2x with identical outputs (~1.6x identical-output speedup for structure prediction models)
- FlashPairformer kernels: 2.7–2.9x faster triangle attention and 1.7–3.2x faster triangle multiplication than the field standard (NVIDIA cuEquivariance / BioNeMo-IR)
- ‘Big’ mode accurately folded human mitochondrial complex I, the TRiC chaperone, a proteasome and a bacterial 70S ribosome (>10,000 tokens) on one GPU node; for comparison, AlphaFold3’s accurate 40S ribosome prediction was 7,663 tokens
- Inference ran on 31,000–70,000+ token viral capsids on one 8-GPU B300 node, but those predictions collapsed (no accurate structure)
- Binder design with one H200, 24 h, a ~1,100-word prompt and no sub-agents matched the in silico ipSAE scores of the earlier campaign at about two orders of magnitude fewer GPU hours (~$150 of GPU plus tokens); tested with Mythos 5.1, Mythos 5 and Opus 5 on 16 targets
- Competition with Adaptyv Bio: five problems, wet-lab validation of 5,000+ designs, up to $1M in Claude credits, $250K in Modal compute credits, DNA from Twist Bioscience
- Same day: Life Sciences Verification Program opened in public beta
What happened
Anthropic’s August binder campaign had let Claude spend up to $10,000 of Modal compute per target, about 2,500 H100 hours. To make the approach affordable, Anthropic had Claude optimize the open-source models it relies on. Claude wrote general kernels (FlashPairformer) for the cubic-cost triangle operations in Pairformer-style models such as AlphaFold3, OpenFold3 and Boltz-2. It also made model-specific changes, such as caching recomputed work and collapsing dead branches. Anthropic checked that downstream accuracy did not change. A low-memory mode extended accurate folding to molecular machines about 1.5 orders of magnitude beyond the models’ training context.
Why it matters
Expert engineers normally need weeks per model to make such optimizations, and the work rarely transfers between models. Here an AI did it across a whole ecosystem of scientific software and released the code. It lowers the compute barrier for academic protein design by roughly 100x. The results are Anthropic’s own benchmarks, and faster models do not add new scientific validation.
Sources
7 sources from 6 sites. Numbers match the chips in the text.
7 sources: 5 primary, 1 press, 1 reaction
Primary
- Anthropic: How Claude is uplifting biomolecular modelinganthropic.com, official
- Technical report: Accelerating open-source biomolecular models with Claude (PDF)www-cdn.anthropic.com, paper
- Code: anthropics/uplifting-biomolecular-modelinggithub.com, code
- Anthropic × Adaptyv 2026 protein design competition (Proteinbase)proteinbase.com, official
- Introducing the Life Sciences Verification Programanthropic.com, official
Press
Reactions
Changes
- Filed (the post had not been cited)