What happened
Divergence Labs’ preprint, released with Anthropic’s 27 October 2026 drug-development update, describes an agentic computational campaign toward a biparatopic VHH that would strip receptor-bound IgE from FcεRI rather than only neutralize circulating IgE, the mechanism of omalizumab (Xolair). Erives, Ferrara, Michael, Mallipeddi, and Jing report that Claude, guided by the research team, decomposed the objective and ran 22 near-autonomous phases, 15 of them computational design. Anthropic says Claude then generated roughly 35,000 antibody designs in about a month; the preprint’s counts are 11,845 scored disruptor redesigns of the known allosteric VHH sdAb026 and 23,587 unique de novo anchor designs (~1,140 and 1,310 H100-hours). Sequences were produced by biomolecular generative models Claude directed (inverse folding, tempered redesign, interface-geometry losses, co-fold critics), not by Claude emitting sequences unaided. Humans set the therapeutic thesis and biparatopic architecture, dropped one redesign scope, constrained linker search, and ran the assays. In Divergence’s laboratory, competitive bio-layer interferometry on soluble FcεRIα ectodomain found 20 of 91 nominated disruptors active. After purification, variants V1, V4, and V8 removed more IgE than omalizumab at every concentration from 25 to 400 nM, but did not improve on parent sdAb026. Three of 57 de novo anchors bound IgE with apparent KD 3.4, 32.5, and 235 nM; affinities may be confounded by avidity. The two arms were not assembled into a biparatopic. No cellular disruption, anaphylactogenicity, animal, or clinical data are reported. Sequences and a public dataset were not released. The disruption mechanism is known: Excellergy has already entered phase 1 on omalizumab-derived disruptors, and sdAb026 is a published Ablynx VHH. This ledger records an experimental hit-generation result, not a food-allergy medicine. Primary sources name Claude, not a SKU; Papex’s report of Claude Mythos 5.1 is not in the preprint or Anthropic post.