Anthropic: Claude Accelerates Over 30 Biological Models by About 4 Times
Anthropic stated in a scientific blog that Claude optimized more than 30 open-source biological molecular models in less than four weeks, achieving an average acceleration of about 4 times, partly by writing custom software for GPUs and optimizing code, all of which is open-source. These models are used for molecular system structure, drug-like molecule design, and gene mutation effect prediction, which were previously costly in inference, limiting laboratory use.
The internal general research model produced 36 optimization packages covering six families: 14 for co-folding and structure prediction, 3 for hallucination, 6 for structure generation, 3 for reverse folding, 7 for genomics, and 3 for protein language models. In precise mode on H100, 14 structure prediction models were on average 1.6 times faster while producing the same output; in fast mode, they averaged 4.1 times faster; and in large molecule mode, they averaged 3.4 times faster. In precise mode, genomics and protein language models were accelerated by about 1.2 to 5 times. Fast mode for structure prediction ranged from 2.3 times for OpenDDE to 6.4 times for Chai-1.
On common bottlenecks, Claude wrote the first version of the FlashPairFormer custom kernel. The triangular attention was 2.7 times faster than the commonly used kernels of AlphaFold3 and Boltz-2 at a pairing width of 128, and 2.9 times faster than ProteNix v2 at a width of 256; triangular multiplication was 1.7 times and 3.2 times faster, respectively. Additional features include intermediate result caching, constant branch folding, CUDA graph capture, and kernel fusion. The low memory mode allows accurate predictions for biological molecular systems (amino acids, nucleotides, small molecules, and ion atoms) with over 10,000 tokens on a single NVIDIA GPU node.
The supervisors were two technical staff from Anthropic who understood biological molecular modeling but had no prior experience in inference optimization or kernel engineering. The company stated that such optimizations typically require a senior team several weeks per model and are difficult to transfer between models. Previously, Mythos 5.1 had accelerated 7 open-source biological models by up to 2.5 times. At the same time, they co-hosted a protein design competition with Adaptyv Bio, offering up to $1 million in Claude credits and conducting wet lab validation for over 5,000 designs.
This treats the general model as a compiler backend: rather than retraining structural networks, it modifies the inference graphs and kernels of the open-source stack. Claude Science has integrated Evo 2, Boltz-2, OpenFold3, and others into its workflow. The open-source optimization packages reduce the compute time for single folding and design, rather than discovering new folding principles.
The market mechanism dilutes the scarcity of GPU hours. The buyers are structural biology and drug design laboratories constrained by inference bills; the sellers are high-performance computing teams writing a set of kernels for each model, with delivery cycles measured in weeks. The beneficiaries are open-source workflows that can run larger systems on a single node and clouds selling H100 hours; the pressured parties are commercial structural prediction services that create usage barriers due to slow inference. Funding has shifted from 'buying another batch of cards for hard runs' to 'running multiple sequences on the same card.'
Source: Public Information
ABAB AI Insight
After AlphaFold proved that structures can be predicted, the bottleneck shifted from algorithms to memory and wall-clock time for each additional atom. Anthropic's ability to modify over thirty repositories in four weeks with a general model signifies that kernel engineering can also be delegated. The fact that two biological staff without CUDA kernel experience could oversee this project indicates that the migration cost has shifted from 'hiring a GPU team' to 'granting model repository access and a single H100.' Open-sourcing all optimization code turns acceleration itself into a customer acquisition strategy, rather than treating kernels as private assets.
The capital path is a commission from scientific workstations. Claude Science sells orchestration and computational scheduling; the faster the model, the more designs can be completed with the same credits, and wet lab competitions further convert credits into real binding data. NVIDIA's BioNemo has already integrated the acceleration library into the same workflow. Money flows from pharmaceutical companies' internal high-performance computing setups to task-based models and cloud quotas. Whoever controls the optimized open-source stack controls which command laboratories default to.
The analogy should be drawn to LLVM and PyTorch compilers. Compilers eliminate handwritten assembly; now language models eliminate attention kernels written per model. DeepMind's open-sourcing of AlphaFold weights changed the discovery rights, while Anthropic's open-sourcing of inference kernels changed the unit cost. The industry's position has shifted from 'is there a structural model' to 'how large a system can be folded on the same card,' marking a phase of toolchain commercialization.
Structural judgment belongs to technological substitution. What is being replaced is the inference engineers assigned per model. The mechanism is that structural prediction networks share abstract bottlenecks like triangular attention; once a general model can write reusable kernels, the marginal optimization cost for edge models approaches that of running another round of delegation. The rate of scientific discovery is beginning to correlate with the unit cost of compute time rather than the number of people who can write CUDA.
ABAB News · Cognitive Laws
The next acceleration of scientific models may come from agents monitored by those who cannot write kernels.
After weights are open-sourced, the next to be open-sourced is the inference kernel.
Running larger molecules on the same card changes laboratory rhythms more than publishing another structural paper.