The Model Ran the Whole Pipeline, Not Just the Sequence Step
The notable part is not the phrase "AI designs proteins" — specialized structure and sequence design models have done that for a while. The difference is the division of labor. Given one detailed human-written prompt, Claude went off and researched the targets, picked epitopes, orchestrated open-source structure and sequence design tools, ran optimization cycles in silico, and screened candidates for solubility, expressibility and novelty, with humans involved only for infrastructure approvals. Anthropic's framing: this is work that takes a human operator weeks. Validation was external. Twist Bioscience received 1,260 of the 1,320 designs (all but the latent GDF-8 set) and, in parallel with Adaptyv Bio, expressed them as human IgG1 Fc fusions. The two CROs return different readouts — Adaptyv classifies designs as binder or non-binder, Twist reports fitted rate constants — so each dataset was labeled independently and blind to the other, then combined under a fixed rule. To rule out the model recycling successes from its training data, the set included two fresh targets from Adaptyv's recent competitions, 15-PGDH and GDF-8. Results varied sharply by target: 72 of 90 designs bound TREM2 and 54 of 90 bound VEGF-A. On RBX1 — a target where an open design contest produced only 9 binders from 245 candidates — Claude got 28 of 90. At least six targets yielded high-affinity binders, and at least four matched or exceeded the best published affinity.
The Analytical Chemistry Numbers Released Alongside
A second experiment used Opus 5, which is generally available. It processed raw NMR data in 23 minutes, with hydrogen counts within 0.08 ¹H of the lab's own reading, and computed LC-MS purity of 96.4% against the lab's 96.33%. This part unlocks no new capability, but it is the more reproducible half of the announcement.
Caveats and the Access Timeline
Pushback has focused on practical value: critics argue the affinities are low for peptide therapeutics and that none of the binders target intracellular proteins. Anthropic also notes that different sources use different statistical approaches, so the various hit-rate figures cannot be substituted for one another. Adaptyv's own assessment is that this was an open-loop experiment — the next step is closing the loop so an agent learns from each experimental batch. The practical line for readers: protein design and related dual-use biology capabilities remain unavailable for general access in Claude Fable 5. Anthropic says an access program for scientists is one of its highest priorities but gave no timeline — you cannot reproduce this pipeline through the API today.
via: Anthropic research report, Anthropic technical paper (PDF), Adaptyv Bio case study, The Decoder