Claude is moving beyond answers and into protein design. Anthropic says its models ran design campaigns on their own, producing working molecules for 14 targets out of 15 while reaching success rates above the industry norm.
That result places Claude inside an early stage of drug discovery, where researchers need molecules that can grip a chosen target. The models achieved success rates between 22% and 35% for molecules that actually gripped their target, compared with the typical industry range of 10% to 15%.
Claude Takes On Protein Design Campaigns
Anthropic’s Mythos Preview and Opus 4.8 models powered the protein-design research. The models ran campaigns largely on their own, and the resulting designs yielded working molecules against 14 of 15 targets.
That 14-out-of-15 figure gives the experiment its clearest signal. Claude did not only produce possible protein designs; the designs led to working molecules across nearly every target included in the campaign. Anthropic also measured how often those molecules gripped their intended target, with results reaching as high as 35%.
The comparison with industry performance makes the result stand out. A typical success rate sits between 10% and 15%, while the Claude models reached 22% to 35%. Those figures show a gap between generating a design and producing a molecule that performs the needed task, and Anthropic says Claude cleared that bar at rates above the norm.
Protein design sits at an important point in early drug discovery because the process must connect a designed molecule with a target. The verified results focus on that connection: whether the resulting molecules worked and whether they gripped the target. Claude’s performance gives Anthropic a concrete result to measure rather than a promise about what AI might do in biology.
Opus 5 Turns Raw Data Into A Faster Readout
A separate test showed another capability. Opus 5 opened raw instrument files without using lab software, then measured a sample at 96.4% pure in 19 minutes.
The lab’s own report took four days. That contrast puts the focus on the time needed to move from raw instrument data to a usable measurement. Opus 5 handled the raw files directly, while the lab’s report arrived after four days.
These results describe two different parts of the research process. Mythos Preview and Opus 4.8 worked on protein-design campaigns and produced molecules that gripped targets. Opus 5 worked with raw instrument files and delivered a purity measurement, connecting model capability to a concrete laboratory result.
The experiments also keep the claims specific. Anthropic has not described Claude as completing every part of drug discovery, and the results do not say that the models replaced a lab. They show success on 14 targets out of 15, target-gripping rates of 22% to 35%, and a 96.4% purity measurement produced in 19 minutes.
Anthropic’s Biology Ambition Comes Into View
Dario Amodei, CEO of Anthropic, wrote that Anthropic hoped for “early glimmers in the coming months” in biology and medicine in an X post last week. The new protein-design and instrument-analysis results provide those glimmers in measurable form.
The timing matters. The results were reported on Aug 20, 2026, giving Anthropic a public milestone for its work with AI models in biology. The company’s research connects model autonomy with laboratory output, from creating candidates to checking a sample’s purity.
Twist Bioscience served as the lab creating candidates, while Adaptyv Bio measured candidates. Their roles frame the results around a chain of work: Claude models generated designs, Twist Bioscience created candidates, and Adaptyv Bio measured candidates. The verified findings do not add a broader claim, but they do show AI activity spanning design and measurement.
What happens next will depend on whether these results hold across more targets and future campaigns, but the current numbers already establish a striking benchmark. Claude reached working molecules on 14 of 15 targets, surpassed the typical 10% to 15% success range with results of 22% to 35%, and Opus 5 cut a raw-file measurement to 19 minutes instead of four days.
Anthropic now has more than an AI model that discusses biology. It has models demonstrating protein design and laboratory data analysis through specific, testable results. The next phase of its biology push will show whether this early momentum can grow beyond these campaigns.
