Claude accelerates open‑source biomolecular models 4× faster and enables >10k‑token predictions on a single GPU

TL;DR

Anthropic released a suite of Claude‑driven optimizations that make over 30 open‑source biomolecular modeling tools roughly four times faster on average, cut memory use enough to predict structures larger than 10,000 tokens on a single NVIDIA GPU, and announced a $1 million protein‑design competition co‑sponsored with Adaptyv Bio.

Claude‑powered inference acceleration

Key result: Claude accelerated 30+ deep‑learning models for structure prediction, protein design, protein language modeling, and genomics, achieving an average 4× speed‑up with minimal precision loss (≈1.6× speed‑up with identical outputs). The work was completed in under four weeks, a task that normally requires weeks of engineering per model.

  • Claude acted as a general‑purpose research model, guided by two Anthropic staff members with domain expertise but no prior kernel‑engineering experience.
  • Optimizations included custom kernels for triangle attention and multiplication (FlashPairformer), caching of redundant work, and pruning of dead code paths.
  • The accelerated models retained downstream performance on tasks such as protein structure prediction.

"Claude accelerated over a dozen biomolecular structure prediction models, achieving, on average, a roughly 4× speed‑up with minimal decrease in precision and a roughly 1.6× speed‑up with identical outputs." – Anthropic blog post

FlashPairformer: custom kernels for triangle operations

Key result: FlashPairformer, a set of Claude‑generated kernels, outperforms the field‑standard cuEquivariance implementation by 2.7–2.9× on triangle attention and 1.7–3.2× on triangle multiplication, depending on model configuration.

  • Triangle attention and multiplication dominate runtime and memory in AlphaFold‑class models because they scale cubically with token count.
  • By rewriting these kernels, Claude achieved state‑of‑the‑art performance for the core Pairformer architecture.

Claude’s FlashPairformer kernel speed‑up over cuEquivariance

Low‑memory “Big” mode for massive biomolecular systems

Key result: Claude introduced a low‑memory “Big” mode that enables accurate inference for systems >10,000 tokens on a single NVIDIA GPU node and successful inference for systems up to 70,000 tokens on one B300 node.

  • Demonstrated accurate predictions for human mitochondrial complex I, the TRiC chaperone, a proteasome, and a bacterial ribosome (each >10,000 tokens).
  • Compared to AlphaFold 3’s 7,663‑token 40S ribosome prediction, Claude’s Big mode pushes the size frontier by ~1.5 orders of magnitude.
  • When pushed to 31,000–70,000‑token viral capsids and protein compartments, predictions collapsed, indicating a lack of generalization far beyond the training context, but the computational barrier was dramatically lowered.

Large molecular complexes folded with Claude’s Big mode

Efficient de novo protein binder design

Key result: Using a single Claude model on one NVIDIA H200 GPU for 24 h, Anthropic reproduced the in‑silico binding scores of its earlier Mythos 5.1 campaigns while spending roughly 100× fewer GPU hours and about $150 in total GPU + token costs.

  • Three Claude variants (Mythos 5.1, Mythos 5, Opus 5) were evaluated on 16 targets.
  • Median and top ipSAE scores matched those of the prior multi‑agent, $10 k‑per‑target experiments.
  • This demonstrates that Claude’s accelerated biomolecular stack can deliver comparable design quality with dramatically reduced compute budgets.

Claude’s ipSAE performance vs. GPU spend

Open‑source release and community competition

  • All optimized code for the 30+ models is available at https://github.com/anthropics/uplifting-biomolecular-modeling.
  • A detailed technical report (PDF) provides benchmarking methodology and full results.
  • Anthropic and Adaptyv Bio are co‑sponsoring a protein‑design competition with up to $1 M in Claude credits, $250 k in Modal compute credits, and wet‑lab validation for >5,000 designs. The competition targets challenging problems such as cross‑species reactivity, pH‑sensitivity, peptide‑MHC specificity, and GPCR design.

"We’re committing up to $1 million in Claude credits and $250,000 in Modal compute credits, as well as wet‑lab validation for over 5,000 designs." – Anthropic announcement

Implications for the life‑science community

  1. Democratization of high‑performance modeling – Researchers can now run state‑of‑the‑art structure prediction and design pipelines on a single GPU, lowering the barrier to entry for labs without large compute clusters.
  2. Accelerated discovery cycles – Faster inference and reduced GPU spend enable more design iterations per project, potentially shortening the timeline from in‑silico concept to experimental validation.
  3. Catalyzing open‑science collaborations – The open‑source release and competition provide a shared benchmark and incentive structure that can align community efforts around reproducible, high‑impact protein engineering.

Further resources

This article faithfully summarizes Anthropic’s September 2026 research announcement without adding or altering any factual content.

Sources