Anthropic published an experiment report on August 18: Claude was tasked with designing small-molecule binders from scratch for a set of protein targets, and it produced usable designs for 14 of 15 targets. Synthesis and testing were carried out by Adaptyv Bio and Twist Bioscience — the model didn't grade its own work.
The numbers carry more weight than any conclusion. In multi-target mode, Mythos Preview posted a 26.7% hit rate and Opus 4.8 came in at 22.6%; switching to single-target focus pushed Mythos Preview up to 35.1%. For comparison, publicly reported hit rates in this field typically run 10-15%. In total, the 1,320 designs yielded 354 experimentally validated binders, with six targets producing high-affinity binders — four of which matched or exceeded published state-of-the-art results.
Some of the Toughest Targets
RBX1 best illustrates the gap. Claude's hit rate on this target was 40%, versus an average of just 3.7% among entrants in a competition on the same target — and the final designs outperformed that year's winning entry.
For TNFα, the work went cross-species: the same batch of designs bound the human, macaque, and mouse versions of the protein simultaneously. That matters a great deal in drug development — whether animal studies and human trials can share a single molecule directly determines how long early-stage development takes.
Another set deliberately went after difficulty: binders with high beta-sheet content. These structures have always been hard for computational design, since beta-strand pairing leaves little room for error. This round produced 15 functional binders with beta-strand content of at least 20%.
Anthropic also wrote up the failures: two targets, BBF-14 and MBP, underperformed. Nobody trusts a report that only lists wins — including these two actually makes the rest of the data more credible.
The Compute Bill
This pipeline isn't cheap. A multi-target run consumes up to 12,500 NVIDIA H100-hours and takes 48 hours of wall-clock time; single-target mode uses 2,500 H100-hours per target over 24 hours. The starting prompt runs about 30,000 tokens.
At current cloud H100 rental rates of roughly $2-3 per GPU-hour, 12,500 GPU-hours works out to $20,000-30,000 per run — spread across 15 targets, that's roughly $2,000 per target. That's a rounding error in an early-drug-discovery budget: a single round of protein synthesis plus affinity testing in the wet lab already costs about the same, and the traditional route needs multiple rounds of trial and error to approach the same hit rate. Compute has become the cheapest link in this chain — a conclusion that wouldn't have held a few years ago.
A Side Project, Done in Passing
The same batch of experiments included a separate analytical-chemistry test using Claude Opus 5. The task was processing real instrument output: it worked through an NMR spectrum in 23 minutes and LC-MS data in 19 minutes. The LC-MS raw file was in a proprietary format, and the model parsed it itself, cross-checking data across 2,664 scans.
On accuracy, hydrogen-atom counts deviated from manual results by no more than 0.08 ¹H, and purity came out at 96.4% against a lab reference value of 96.33%. After finishing, the model independently proposed which follow-up validation experiments to run next — and its plan matched what the lab had already scheduled.
This part may get overshadowed by the protein numbers above, but it shouldn't be. The most labor-intensive part of drug discovery isn't the design stage — it's interpreting and filing dozens to hundreds of spectra a day, which is a postdoc's daily grind. A model that can read proprietary instrument file formats and decide what to do next on its own is chipping away at exactly that.
A Deliberate Limit
Anthropic states plainly in the report that protein design and other dual-use biological research capabilities are not available to general users on Claude Fable 5. That's a gate the company set deliberately: the same capability that designs binders could be used to design other things, and Anthropic chose to keep it in a controlled channel rather than ship it straight into a consumer product.
The datasets and prompts used in the experiments are already up on Hugging Face, with separate technical reports published for the protein design and chemical analysis work. Releasing the prompts alongside everything else suggests Anthropic sees the real barrier to this method as compute and wet-lab work, not prompt engineering itself.
Sources: Anthropic's official research page, Adaptyv Bio and Twist Bioscience experiment records, CocoLoop, Hugging Face public datasets; hit-rate definitions, H100-hour figures, and the 96.4% purity reference value were cross-checked.