In September 2026, Anthropic published a research report titled *How Claude is uplifting biomolecular modeling* (September 17, 2026), describing how Claude wrote a set of custom GPU kernels — codenamed FlashPairformer — targeting triangle attention and triangle multiplication, the two operations in structure-prediction models that Anthropic described as "cubic in both runtime and memory." Doubling system scale costs 8x in time and memory; tripling costs 27x.
What the 36 toolkits contain
The deliverables are not a model but 36 "drop-in" optimization kits covering more than 30 open-source models. The repository is anthropics/uplifting-biomolecular-modeling on GitHub, under Apache 2.0. Each kit places the upstream original and optimized versions side by side, sharing a runtime that includes FlashPairformer and the triangle-attention kernels.
| Category | Kits | Representative models | Architecture | |---|---|---|---| | Co-folding / structure prediction | 14 | AlphaFold3 (JAX and Torch), OpenFold3, Boltz-2, Protenix, Chai-1, ColabFold | Diffusion, flow matching | | Structure generation | 6 | RFdiffusion3 et al. | Diffusion | | Genomics | 7 | Borzoi et al. | Convolution / attention | | Hallucination-based design | 3 | Not enumerated | Graph neural networks | | Inverse folding | 3 | Not enumerated | Graph neural networks | | Protein language models | 3 | ESM C et al. | Transformer |
Three modes for three real use cases
Each kit preserves original model weights and offers three modes:
| Mode | Models tested | Avg. speedup on H100 | Accuracy note | |---|---|---|---| | Exact | 14 | 1.6x | Bit-identical to original output | | Fast | 13 | 4.1x | Minimal numerical differences allowed | | Big | 13 | 3.4x | Trades memory for scale |
Within Fast mode, the fastest was Chai-1 at 6.4x; the slowest, AtlasFold, at 2.8x. Exact mode never exceeded 2.5x on any model.
FlashPairformer's kernel-level gains: triangle attention was 2.7x faster than domain-standard kernels at pair width 128 (AlphaFold3/Boltz-2 config) and 2.9x at pair width 256 (Protenix v2); triangle multiplication was 1.7x and 3.2x respectively. At 2,048 tokens with pair width 128, gains reach 4x. Notably, the "domain-standard" baselines are NVIDIA's own specialized kernels — first cuEquivariance, then BioNeMo Inference Runtime — so these multipliers should not be read as gains over unoptimized code.
Did accuracy drop?
The report's yardsticks: DockQ (above 0.23 counts as acceptable) for interface prediction, and ipSAE for design results. On FoldBench-Lite (1,925 model–target pairs across 13 model configurations), the fraction of acceptable interfaces was: default 54.8%, Exact 55.0%, Fast 54.5%, Big 54.2% — "no discernible pooled change" per Anthropic. The pooled statistic averages out per-model differences, and the report does not give per-model accuracy distributions.
How far Big mode reaches — and where it stops
Big mode buys scale, not speed. AlphaFold3's flagship system in the original paper was the 40S ribosome at 7,663 tokens; training slices were capped at ≤768 tokens. Big mode can accurately model assemblies over 10,000 tokens on a single node, and run inference past 70,000 tokens.
Folded large assemblies — human mitochondrial complex I, the TRiC chaperonin complex, the proteasome, and the bacterial 70S ribosome — achieved TM-scores between 0.92 and 0.997 versus experimental structures ("to our knowledge," per the report).
The boundary appears at larger scales. Using a single node of 8x B300, the team ran seven structure predictions on assemblies of 31,000–70,000 residues (virus capsids, protein compartments). All seven completed; all seven collapsed into a compact sphere, with scoreable TM-scores of 0.08–0.14. The report attributes this to generalization failure: these assemblies are 40–90x larger than AlphaFold3's largest training slice. The report frames this as "extending what is computable, not extending what the model has learned."
From $10,000 per target to $150
For binder design, this round differed sharply from the April 2026 round:
| | April 2026 round | This round | |---|---|---| | Prompt length | ~16,000 words | ~1,100 words | | Sub-agents | Yes | No | | Human intervention | Yes | No | | Compute | Up to $10,000/target, ~2,500 H100-hours | 1x H200 for 24 wall-clock hours | | Total cost | ~$10,000 per target | ~$150 (GPU + tokens) |
In silico scores: Opus 5 median ipSAE 0.785, Mythos 5.1 0.781, Mythos 5 0.739 (vs. 0.749 in the earlier round); best designs scored 0.833, 0.825, 0.813 (vs. 0.817). "Comparable," not "better," is accurate — Mythos 5's median is lower. These are computational predictions, not wet-lab measurements; the designs are experimentally unverified. The follow-up is a protein-design competition with Adaptyv Bio: five targets (including species cross-reactivity and difficult GPCR targets), 5,000+ designs to be validated experimentally, with up to $1M in Claude credits from Anthropic, $250K in compute from Modal, and DNA synthesis from Twist Bioscience.
Who did the work, and who checked it
Code optimization was primarily done by Claude within the Claude Science environment. The two Anthropic technical staff supervising it were familiar with biomolecular modeling but had no prior inference-optimization or GPU-kernel experience. Models tested included Mythos 5.1, Mythos 5, and Opus 5. An earlier public result had Claude Mythos 5.1 accelerate 7 open-source biology models by up to 2.5x; this round extended coverage to 30+ models.
What was and wasn't controlled
The report's own stated limitations:
- Measurements only on H100 hardware
- ColabFold comparison used version 1.6.1; that project has since released its own fused kernels
- Exact and Fast modes use up to 3.2x more memory
- No third-party independent testing of FlashPairformer
- The repository is a "reference release" — not actively maintained, not accepting contributions (code is free to fork, but PRs will not be merged)
- Design results not experimentally validated
Where this sits
The most direct significance may not be in biology but in the next stage of AI coding. Typical coding-agent benchmarks (issue fixing, refactoring, frontend work) share one trait: verification is cheap — run the tests. FlashPairformer-style work is different: verification is extremely expensive, requiring knowledge of triangle-attention memory access patterns, DockQ thresholds, and pooled benchmark interpretation. Whether AI-written scientific code is usable depends on whether someone can build a sufficiently rigorous external control.
The same week, a parallel path appeared: on September 17, AItonomy Foundation released open-source ScienceIDE, sponsored by PhAI Labs and Alibaba Qwen, aiming to convert the world's scientific codebases into executable, verifiable agent learning environments, and trained PhAI-IDE-72B / 9B / 4B models on verified interaction trajectories (arXiv:2609.19134).
The two paths ask opposite ends of one question. Anthropic's asks: can AI make existing scientific software faster? ScienceIDE asks: can that capability be distilled and transferred to other research domains? What likely decides success is labor: how much anyone is willing to pay for verification. AlphaFold3's training slice cap was 768 tokens; Big mode ran past 70,000 tokens yet collapsed into a sphere. The distance between those two numbers is the distance between "computable" and "computed correctly." Who will measure it?
---
References
1. Anthropic, *How Claude is uplifting biomolecular modeling*, 2026-09-17. https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling 2. Anthropic, *anthropics/uplifting-biomolecular-modeling* (GitHub, Apache 2.0, 36 optimization kits), 2026-09-17. https://github.com/anthropics/uplifting-biomolecular-modeling 3. Unite.AI, *Anthropic Reports Claude Optimized 30+ Open-Source Biomolecular Models*, 2026-09-18. https://www.unite.ai/anthropic-reports-claude-optimized-30-plus-open-source-biomolecular-models 4. AI Insiders, *Anthropic says Claude rewrote 30 biology models to run 4x faster*, 2026-09-18. https://aiinsiders.net/article/anthropic-says-claude-rewrote-30-biology-models-to-run-4x 5. AItonomy Foundation, *ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments*, arXiv:2609.19134, 2026-09-16.