Anthropic disclosed on August 18, 2026, that its Claude models had completed the first fully autonomous, end-to-end de novo protein binder design campaign and submitted the results to outside laboratories for physical confirmation. External evaluators Adaptyv Bio and Twist Bioscience independently produced and tested Claude's designs in the lab, finding that of the 15 targets the system designed against, it successfully produced binders against 14 of them. In a multi-arm campaign, Claude's Mythos Preview and Opus 4.8 models achieved overall hit rates of 26.7% and 22.6%, respectively, during a 48-hour session; when Mythos Preview focused on individual targets in 24-hour sessions, the hit rate rose to 35.1%, compared with the 10 to 15% typical in current protein design campaigns. Designing a new protein binder from scratch, known as de novo design, has historically taken protein engineers months per target.
The targets spanned some of medicine's most-studied problems. They included PD-L1, a checkpoint protein central to cancer immunotherapy; TREM2, implicated in Alzheimer's disease; TNFalpha, the target of blockbuster anti-inflammatory drugs including Humira; and EGFR, a well-established oncology target. Performance varied sharply by target. Against TREM2, 72 of 90 Claude-designed proteins bound, an 80% hit rate on a target directly relevant to Alzheimer's disease research. Against RBX1, the contrast with human competition entrants was starker still: the same target had been the subject of an open design competition run by Adaptyv Bio; Anthropic had the competition's winning design physically reproduced and tested on the same assay plate as Claude's top candidate, which bound at approximately 3.9 nM, roughly ten times more tightly than the competition winner's approximately 45 nM. Adaptyv Bio's own case study noted that Claude's designs had noticeably higher success rates than competition entrants and would have won five of six past design competitions, yielding tighter binders. The pipeline ran largely autonomously: Claude researched targets, selected binding sites, orchestrated open-source structure and sequencing design tools, optimized candidates in silico, and generated ranked designs while screening for solubility, expressibility, and novelty.
The results arrived with significant caveats, both scientific and structural. The system did not work everywhere; against maltose-binding protein, a notoriously smooth target, none of the 90 designs was confirmed to bind. Anthropic acknowledged that "protein binders are not drugs" and that creating a high-affinity binder is only the first step in drug development. Access to the capability itself remains restricted: protein design and other dual-use biology capabilities are currently blocked for general access in the highest-performing models because Anthropic has determined it cannot yet reliably distinguish legitimate drug discovery queries from potential bioweapon development applications. On the scientific side, Martin Shkreli, a former pharmaceutical executive, called the work "not impressive" and criticized the low affinities of the binders, noting that none of Claude's designs targeted intracellular proteins. Anthropic itself states it plans further, more extensive characterization to confirm these hit rates and affinity measurements, and the current figures should be treated as first-round results pending that follow-up.
The structural bottleneck the experiment exposed is as consequential as its headline numbers. The experiment confirmed that while design software is free, the significant expense and bottleneck now lie in the physical synthesis and experimental validation of proteins, underscoring the scarcity of wet-lab capacity. The multi-target runs consumed up to 12,500 hours on Nvidia H100 chips. Adaptyv Bio described this campaign as an open-loop experiment in which Claude designed proteins and Adaptyv tested them; the next step, it said, is to close that loop so an agent proposes a batch, receives experimental data, and chooses the next batch based on what it just learned.
The clearest takeaway is not that AI has solved drug discovery, but that the computational step, once the rate-limiting constraint, has been largely removed, and the queue now forms at the wet lab.