Anthropic's Claude Speeds Up 36 Biology AI Tools by 4x
Anthropic used its Claude AI to optimize 36 open-source biomolecular modeling tools, achieving a 4x average speedup that dramatically lowers the cost of computational biology.
Anthropic has released Apache 2.0-licensed optimization code for 36 open-source biomolecular modeling tools, including AlphaFold3, OpenFold3, Boltz-2, Chai-1, Protenix, RoseTTAFold3, and several others like RFdiffusion and ESM-C. Supervised by two staff members with no prior GPU kernel engineering experience, the Claude model completed the optimization work in under four weeks. The resulting code delivers an average 4x acceleration in fast modes and a 1.6x speedup for bit-identical outputs.
At the core of these improvements is FlashPairformer, a custom GPU kernel written with Claude's assistance to address bottlenecks in triangle attention and triangle multiplication. FlashPairformer outperformed NVIDIA's cuEquivariance and BioNeMo Inference Runtime, yielding 2.7x to 2.9x acceleration on triangle attention and 1.7x to 3.2x on triangle multiplication. Additionally, a low-memory Big mode allows researchers to model massive biomolecular complexes exceeding 10,000 tokens, such as a bacterial ribosome, on a single GPU node. Anthropic even successfully processed inputs above 70,000 tokens on an eight-GPU B300 node, though very large viral capsids collapsed due to training-context limits.
For practitioners, these optimizations translate to massive cost savings. Anthropic rebuilt its agentic binder-design pipeline using the optimized models, matching the performance of previous campaigns while using roughly 100x fewer GPU hours. The new setup cost just $150 in combined GPU and token spending, compared to the previous cost of up to $10,000 or 2,500 H100 GPU-hours per target. To encourage adoption, Anthropic and Adaptyv Bio are co-sponsoring a protein design competition offering $1 million in Claude credits, $250,000 in Modal compute, and wet-lab validation for over 5,000 designs.
The optimization kits support standard environments like Docker, Apptainer, or Python virtual environments, targeting hardware configurations like the H100, A100, H200, B200, and B300. However, Anthropic has explicitly labeled the repository as unmaintained, meaning no pull requests will be accepted and developers must fork the code for long-term maintenance.
This is our own summary of reporting by AlphaSignal
