![\"Anthropic](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/08\/1787490669_623_0x0.jpg\")
<\/p>\n<\/p>\n
A smartphone displaying the Anthropic Claude Science logo is held in front of an illustration of a DNA double helix on a dark blue background in this photo illustration taken in Italy on July 9, 2026. Claude Science, launched by Anthropic on June 30, 2026, is an AI workbench designed to support computational research, including genomics and drug discovery, through a coordinating AI agent with more than 60 curated skills and connectors, according to the company. (Photo by Matteo Della Torre\/NurPhoto via Getty Images)<\/p>\n
NurPhoto via Getty Images<\/p>\n
Anthropic handed Claude sixteen protein targets, a cloud computing account with a hard time limit, and a written protocol. Then it stayed out of the decisions. Over sessions lasting one to two days, the model read up on each target and chose which surface of it to aim at. It installed and ran the design software itself, then delivered thirty candidate proteins per target, ranked best to worst.<\/p>\n
Two outside laboratories then built all 1,320 of those designs exactly as delivered and measured whether they stuck.<\/p>\n
Three hundred and fifty-four of them did, a hit rate of 27%, and among the designs Claude ranked first for each target, 49% bound<\/a>. Fourteen of the fifteen targets that gave readable measurements yielded at least one working design.<\/p>\n Every piece of design software Claude used is open source and free to download. Anthropic paid nothing to license any of it. The technical report<\/a> is specific about this, noting that AlphaFold 3’s weights and two other well-known packages were left out precisely because their licenses were restrictive.<\/p>\n The expense in this experiment sat somewhere else entirely, with the two laboratories that had to physically build 1,320 proteins and find out.<\/p>\n What The Campaigns Did<\/p>\n A binder is a small protein built to latch onto one specific spot on a larger one, which is how a great many drugs work. Designing them has historically taken a specialist weeks to months per target. Most of that time went into judgment rather than code. The specialist had to choose which region of the target to model, which tools to use, how hard to filter, and which thirty designs out of thousands to order.<\/p>\n Anthropic wrote that judgment into a single protocol prompt of roughly 16,000 words and handed it over. Humans picked the targets, paid for the compute, placed the orders and read the final data. Everything in between was the model’s call.<\/p>\n The clearest comparison in the report is a target called RBX1, which had recently been the subject of an open design competition<\/a>. Among the 245 entries humans submitted, nine bound. Of Claude’s ninety designs, twenty-eight bound, and its best one held roughly ten times more tightly than the competition’s winning entry when both were re-made and measured side by side on the same plate. On another target, TNF\u03b1, it produced twelve working designs where several published efforts had reported none.<\/p>\n It also failed cleanly. Against a protein called MBP, none of ninety designs bound.<\/p>\n The Part That Stayed Expensive<\/p>\n The money went to the laboratories. Adaptyv Bio, in Lausanne, received all 1,320 designs, built each one in a cell-free chemical system, and read the binding on an optical sensor across five concentrations. Twist Bioscience, in South San Francisco, received 1,260 of them, grew them as antibody fusions in human cells instead, and ran them on a different sensor format entirely. Neither lab saw the other’s data, or which model or campaign a sequence came from. Both were paid contractors, and the report discloses it.<\/p>\n That is two independent physical builds of more than a thousand proteins each, in two different formats, on two continents. There is no way to simulate it, and the report is blunt about why.<\/p>\n Experimental screening remains the only way to learn which targets a campaign has succeeded on and the affinity of the designed binders.<\/p>\n The scoring models Claude used to rank its own work were good at separating winners from losers within a target, and useless at predicting which targets would fail. Designs against the three targets that flopped scored nearly as well as designs against the ones that produced binders in quantity. Confidence in the model told you almost nothing until a lab checked.<\/p>\n Where Scarcity Moves<\/p>\n The report’s closing suggestion is that campaigns like this “should be within reach of laboratories that have targets of interest but no expertise in computational protein design.” The prompts and data are published on Hugging Face<\/a> for anyone to rerun.<\/p>\n Take that seriously and the consequence is a great many more designed proteins, coming from labs that could never previously staff the work. Every one of those campaigns ends the same way it did here, with an order sent to somebody who owns physical capacity to synthesize and measure. Cheaper design multiplies the work waiting downstream of it.<\/p>\n Two inputs stayed genuinely scarce in this experiment. One is compute: the multi-target runs consumed up to 12,500 hours on Nvidia H100 chips<\/a>. The other is wet-lab throughput, and it is the one almost nobody is pricing.<\/p>\n The companies sitting in that layer are the ones this pattern runs through. Twist Bioscience manufactures synthetic DNA by writing it onto silicon chips and now reports its DNA synthesis and protein work as one combined business<\/a>. Adaptyv Bio has built what it calls a cloud lab for protein designers<\/a> and says it has tested more than 10,000 proteins for over thirty partners. Neither is a bet on which AI model wins, because both get paid regardless of which one sends the order.<\/p>\n What Would Confirm It<\/p>\n The honest limits matter, and the authors state them first. The evidence here is binding and nothing more. No design was tested for whether it does anything useful in a living system, and none had its structure solved, so every picture in the paper is a prediction.<\/p>\n The comparisons carry caveats too. Four of the six competitions Claude was measured against had already published their results where the model could read them, and most of the targets are heavily documented in the literature. The authors also decline to claim Claude beat a human expert, because they never ran the matched comparison.<\/p>\n None of that touches the economics. Whether the design step was done brilliantly or merely competently, it was done with free software on rented compute, and it still ended in an invoice from two laboratories.<\/p>\n Order volume in the synthesis and testing layer over the next few quarters will show whether this is a real shift or one striking result. So will the number of outside labs that bother to rerun the protocol now that it costs nothing to try. Design capacity can now be added with a prompt and a compute budget. Buildings, robots and trained hands still take years.<\/p>\n","protected":false},"excerpt":{"rendered":"A smartphone displaying the Anthropic Claude Science logo is held in front of an illustration of a DNA…\n","protected":false},"author":2,"featured_media":148667,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[4302,53,3154,72868,72869,182,38051,47995,72867,64139,4653],"class_list":["post-148666","post","type-post","status-publish","format-standard","has-post-thumbnail","category-anthropic","tag-ai-drug-discovery","tag-anthropic","tag-anthropic-claude","tag-biotech-ai","tag-biotech-stocks","tag-claude","tag-claude-fable-5","tag-claude-science","tag-de-novo-protein","tag-llm-benchmark","tag-twist-bioscience"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/148666","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=148666"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/148666\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/148667"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=148666"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=148666"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=148666"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}