Anthropic's protein run is checkable, which is rarer than the 26.8% hit rate

Anthropic described a set of autonomous protein binder campaigns in a post on X on Tuesday and published a 29-page technical report ^[1], then released the prompts, predicted structures, provenance records and experimental measurements for all 1,440 designs on Hugging Face ^[2]. That second act is the one operators should care about: a hit rate is a marketing artifact until someone outside the lab can recount it, and this one can be recounted.

The setup is legible. Amir Shanehsazzadeh, the report's sole listed author, wrote a roughly 16,000-word protocol prompt covering the scientific and operational knowledge needed to run a binder campaign, and Claude made the individual design decisions ^[3]. Anthropic ran Claude Opus 4.8 and Mythos Preview as agents against 16 targets; the models researched each target, picked regions and binding sites, installed open-source design and structure-prediction software, generated and computationally optimized candidates, and returned 30 ranked designs per target ^[4]. No epitope, scaffold or sequence was supplied in advance ^[5]. Single-target runs got 24 hours and $10,000 of compute; multi-target runs got 48 hours and $50,000 ^[6]. Anthropic chose the targets, supplied the GPU accounts, set the budgets, ordered synthesis and interpreted the results afterwards ^[7]. Notably, Anthropic introduced no specialized protein-generation model; Claude coordinated existing open-source backbone generators, sequence-design tools and structure predictors ^[8].

Adaptyv Bio and Twist Bioscience, both paid contract research organizations, synthesized and measured the designs ^[9]. One target, mature GDF-8, aggregated during testing, leaving 1,320 designs across 15 targets with interpretable measurements ^[10]. Of those, 354 bound, a reported 26.8% ^[11], which recomputes to 26.8% ^[21]. Claude produced at least one binder against 14 of the 15 interpretable targets ^[12], and 49% of designs ranked first in their campaign bound ^[13].

The averages hide most of the signal. TREM2 returned 72 binders from 90 designs and VEGF-A 54 from 90 ^[14], or 80% and 60% ^[22]. The synthetic beta-barrel BBF-14 produced three weak binders, and maltose-binding protein produced none from 90 designs ^[15]. Anthropic reports that the computational confidence scores gave little warning those campaigns would fail ^[16], which is the finding a buyer should weigh most heavily.

Here is where the released files earn their keep. The summary numbers do not fully reconcile on their own: 1,440 designs minus 1,320 interpretable leaves 120 excluded for a single target ^[23], while the per-target counts quoted are 90 ^[24], and 15 targets at 90 designs each would be 1,350, or 30 more than the stated 1,320 ^[25]. With the dataset published, that is an arithmetic question anyone can settle. Without it, it would be a footnote nobody could audit.

The model comparison is similarly hedged in the source itself. Mythos Preview's single-target campaigns hit 35.1%, against 26.7% for its multi-target run and 22.6% for Opus 4.8 multi-target ^[17], a spread of 12.5 points ^[26]. The single-target format also received 2.8 times the compute per target, and Anthropic says the experiment cannot isolate whether focus, budget or model behavior caused the gap ^[18]. The sharpest external comparison is RBX1, where 28 of Claude's 90 designs bound against nine of 245 entries in a recent open competition ^[19]: 31.1% versus 3.7% ^[27].

Watch whether anyone outside Anthropic recounts the Hugging Face measurements and publishes a diff against the report. Watch for a compute-matched single-versus-multi-target run that answers the question this one cannot. And watch whether the maltose-binding protein and BBF-14 failures get a mechanistic explanation, since confidence scores did not predict them ^[16]. Anthropic also opened Claude Cowork to every paid plan earlier the same Tuesday, according to RuntimeWire ^[20]; the orchestration pattern is the product, and the protein work is its most auditable demonstration so far.