What Claude was given, and what it did with it
Anthropic ran the campaign inside Claude Science, giving the model GPUs for running protein design and folding models, internet access, a corpus of protein-design literature, and connectors to ordinary work tools. Compute was budgeted: up to 12,500 NVIDIA H100 hours in multi-target mode and 2,500 per target in single-target mode. The instruction was to design 30 protein binders per target. [1]
Claude did not invent the design method. It orchestrated publicly available structure design, sequence design and co-folding models, choosing which to use, in what order, and how to run optimization rounds over their output. Anthropic states that beyond delivering the initial prompt, its only involvement was granting access approvals such as network requests and watching the infrastructure, with no additional scientific, technical or operational guidance. [1]
Atlas interpretation: That is the actual claim, and it is narrower and more interesting than "AI designed proteins." The specialist models were already public and already good. What was missing was the ability to wield them: knowing which tool answers which question, reading the failure, and running the next round. Adaptyv makes the same point from the other side, noting that before general models, standing up a protein structure predictor could cost hours of debugging opaque environment errors. The result is evidence about orchestration, not about a new design algorithm. [1][3]
The numbers, and what they are measured against
Across 1,320 designs, 354 bound their target, covering 14 of 15 targets. By mode and model, Anthropic reports a 26.7% hit rate for Mythos Preview and 22.6% for Opus 4.8 in multi-target mode over 48 hours, rising to 35.1% for Mythos Preview in single-target mode with 24 hours per target. It puts the usual rate in protein-design campaigns at 10 to 15%. [1]
Adaptyv Bio’s own account gives an overall hit rate of 26.8% and adds the step before it: 95% of the designs expressed, matching the best expression rates from its EGFR competition among hundreds of expert designers. Sixteen targets were chosen; one, GDF-8 mature, was excluded for low-quality measurements attributed to the target aggregating, and one, maltose-binding protein, produced no binders at all. [3]
Against Adaptyv’s past public competitions, restricted for fairness to de novo minibinders, Claude’s designs would have won five of six and produced tighter binders in the same five. On TREM2 it reached 80% against the 38.3% Adaptyv reported in competition; on RBX1, 40% against 3.7%; on 15-PGDH it improved the best known affinity from 1.7 micromolar to 33.4 nanomolar. It did not beat the best binder from the Nipah virus competition. [3]
Who did the measuring
Adaptyv Bio and Twist Bioscience produced and tested the designs independently. Adaptyv received the sequences anonymized, with no information about which Claude model produced which design, and ran affinity characterization by surface plasmon resonance at five target concentrations with duplicate measurements. [1][3]
Atlas interpretation: This is the part that separates the campaign from most model-capability announcements. The claim is falsifiable by a party with its own reputation at stake, on instruments the claimant does not control, against a benchmark the measuring party had already run with human experts. Anthropic did not grade its own homework, and blinding the designs removes the most obvious way the grading could have drifted. [3][1]
Atlas interpretation: The caveat that survives blinding is target selection. The 16 targets came from Adaptyv’s own previous public competitions, hackathons and BenchBB, which means published designs, published results and published discussion existed for them, and Claude had internet access. Prior exposure does not explain the size of the gaps, an 80% hit rate on TREM2 against 38.3%, or a best 15-PGDH binder roughly fifty times tighter than the previous one, but it does mean the comparison is against entrants who were solving these targets fresh, and it is the first thing an independent replication should vary. [3]
What a binder is not
Adaptyv is explicit that the proteins tested here are not therapeutics and completed only the first step: demonstrating that they bind. The rungs after it are therapeutic formats and developability, function in living cells, organoids and more realistic models, and finally in-vivo evidence. Each requires better models, more compute and more wet-lab data than the one before. [3]
The campaign was also open-loop. Claude designed, Adaptyv tested, and the results were published; the model did not see its own experimental results and choose a second batch on the strength of them. Anthropic says it intends to follow with more extensive characterization to confirm the hit rates and affinity measurements, and reports that Claude struggled against a beta-barrel target as well as maltose-binding protein, where none of 90 designs bound reproducibly. [1][3]
Atlas interpretation: The accompanying analytical chemistry result is smaller and easier to check, which is why it is worth keeping next to the headline. Given raw NMR and mass-spectrometry instrument files, Claude Opus 5 worked out the encoding, reproduced the instrument’s own recorded totals for all 2,664 scans before interpreting anything, reported purity of 96.4% against the lab’s 96.33%, proposed a standard control on its own, and corrected an overstatement in its first reading. That is not a discovery. It is a day of routine laboratory work done in under half an hour, and routine work is most of what a lab does. [1]
Sources
- How Claude is accelerating protein design and analytical chemistry
Anthropic · Aug 18, 2026
- Anthropic says Claude designed working protein binders, and beat human experts on some
The Next Web · Aug 19, 2026
- Case study: Benchmarking Claude's protein designs in the wet lab
Adaptyv Bio · Aug 20, 2026