Anthropic announced that Claude optimized more than 30 open-source biomolecular models in under four weeks, achieving roughly 4x average speedups and enabling researchers to fold massive protein structures previously out of reach on a single graphics processing unit (GPU). The company open-sourced the optimized code and launched a $2 million protein design competition with Adaptyv Bio, signaling a shift toward making cutting-edge molecular biology tools more accessible and efficient.

What Did Claude Actually Accomplish With These Protein Models?

Claude's optimization work covered 36 model implementations across six families of biomolecular tools, including structure prediction, protein design, genomics, and protein language models. The optimization was supervised by just two members of Anthropic's technical staff who had experience in biomolecular modeling but no background in the specialized field of inference optimization or GPU kernel engineering. This constraint makes the results particularly noteworthy: a general-purpose AI model achieved significant speedups without domain-specific expertise guiding every step.

The optimized packages come in three modes designed for different use cases. The "Exact" mode reproduces the original model's outputs bit-for-bit, the "Fast" mode trades a small amount of numerical precision for speed, and the "Big" mode reduces memory consumption so larger inputs fit on a single GPU. On NVIDIA H100 GPUs, the Fast mode achieved an average 4.1x speedup across 13 structure prediction models, while the Big mode averaged 3.4x faster performance.

To achieve these gains, Claude developed custom GPU kernels called FlashPairformer v1, which handle triangle attention and triangle multiplication operations that consume most of the runtime in modern structure prediction models like AlphaFold3 and Boltz-2. These custom kernels ran triangle attention 2.7x to 2.9x faster than industry-standard kernels, depending on the configuration, and triangle multiplication 1.7x to 3.2x faster on a single H100 GPU.

Why Can Claude Now Fold Massive Protein Structures That Were Previously Impossible?

The Big mode optimization unlocked a major capability: accurate modeling of biomolecular systems larger than 10,000 amino acids and nucleotides on a single GPU node, with inference possible on systems larger than 70,000 residues. This is a significant breakthrough because structure prediction models are typically trained on much smaller sequences. AlphaFold3, one of the field's leading models, was trained on crops of at most 768 tokens, making predictions on much larger structures a genuine frontier challenge.

To demonstrate this capability, Anthropic used Claude to accurately fold some of the largest biological structures ever predicted by AI, including the human mitochondrial complex I, the TRiC chaperone complex, a proteasome, and a bacterial 70S ribosome. These predictions matched experimental structures with TM-scores (a measure of structural similarity) ranging from 0.92 to 0.997, indicating near-perfect accuracy. The AlphaFold3 paper highlighted an accurate 40S ribosome prediction of 7,663 tokens as a major achievement; Claude's optimizations now enable predictions on structures 10 times larger.

How to Access and Use Claude's Optimized Biomolecular Tools

• Open-Source Release: Anthropic released 36 drop-in optimization kits on GitHub under the Apache License 2.0, with each kit corresponding to one upstream tool, allowing researchers to integrate optimizations into existing workflows without rewriting code.
• Multiple GPU Configurations: The initial release is configured for NVIDIA H100 80 GB GPUs on Linux x86-64, with additional configurations for A100, H200, B200, and B300 GPUs, ensuring compatibility across different research environments.
• Docker and Container Support: Prebuilt Docker and Apptainer container images are planned within a week of the initial release, simplifying deployment for researchers without deep infrastructure expertise.
• Three Operating Modes: Users can choose Exact mode for bit-perfect reproduction, Fast mode for maximum speed with minimal precision loss, or Big mode for memory efficiency on large structures.

The GitHub repository notes that this is a reference release not actively maintained and not accepting contributions, but the open-source nature means researchers can fork and adapt the code for their own needs.

What's the $2 Million Protein Design Competition About?

Anthropic and Adaptyv Bio jointly announced a protein design competition with $2 million in experimental validation funding. The competition targets five problems at the frontier of protein design, including species cross-reactivity, pH-sensitivity, peptide-MHC specificity, and difficult targets like GPCRs (G-protein coupled receptors). Adaptyv Bio will test more than 5,000 protein designs in its automated laboratory at no cost to participants, while Anthropic provides an additional $1 million in Claude credits.

Three participation tracks are available: research labs and companies can apply for up to $50,000 in Claude credits for academic teams or $25,000 for industry teams, with roughly 20 teams selected for reserved lab testing of up to 18 designs per challenge; individual researchers or small teams of up to three can access Claude Max 20x at no cost during the competition; and a self-supported open track allows optional Claude use. Applications run from September 16 to September 24, 2026, with selected teams notified before September 28, 2026. Five weekly challenges run from September 28 to October 31, 2026.

How Efficient Is Claude's Protein Design Compared to Earlier Approaches?

Anthropic's earlier protein design work required a roughly 16,000-word prompt, multiple sub-agents, and up to $10,000 per target in compute costs, consuming approximately 2,500 H100 GPU hours within 24 hours. The new optimized approach uses a single Claude model with one NVIDIA H200 GPU and 24 hours of wall time, working from a prompt of about 1,100 words with no sub-agents and no human steering. This represents a roughly 100-fold reduction in GPU budget.

On the ipSAE benchmark, an in silico score predictive of experimental binding, the new approach achieved median scores of 0.785 for Opus 5, 0.781 for Mythos 5.1, and 0.739 for Mythos 5, compared to 0.749 for the earlier campaigns. Best-design scores reached 0.833, 0.825, and 0.813 versus 0.817 previously. The combined spend of approximately $150 on GPUs and tokens achieved in silico performance matching campaigns that cost 100 times more, though Anthropic notes these designs have not been tested experimentally and the scores are computational predictions rather than real-world results.

This efficiency gain suggests that as AI models improve and optimization techniques mature, the barrier to entry for protein design research may lower significantly, potentially democratizing access to tools previously available only to well-funded institutions and companies.