The number worth keeping is 17 out of 211. Anthropic said on October 21 that an internal Claude model designed protein minibinders aimed at scorpion neurotoxins, and that seventeen of the ones sent to a lab actually bound.
What the model did
This was not a chat transcript about biology. Anthropic described a research model working for about 64 hours, mostly on its own, from a single prompt and four follow-ups. The output was a set of designed proteins, minibinders, meant to grab onto scorpion neurotoxins. Designing a sequence and folding a picture of it is the cheap part. The expensive part is whether the protein exists in a tube and sticks to the right target.
Adaptyv Bio and Twist Bioscience ran that test on 211 designs. Seventeen cleared the binding criteria the team had set. Anthropic said the full set, hits and misses, is being released with the data. Publishing the misses is what makes the seventeen readable. A highlight reel of winners would have told you nothing about the rate.
What it is not
It is not an antivenom you can buy, and it is not a clinical trial. A binder that works in an assay can still fail in a body, in manufacturing, or on cost. The honest claim is already interesting without those leaps: a model proposed hundreds of candidates, and a normal lab found a small set that did the biochemical job.
If you work in a lab, the release of all 211 designs is the part to download. The rate is the part to argue about. Seventeen is not most. It is also not zero.