Anthropic Left Claude Alone With 15 Protein Targets for 48 Hours, and the Lab Results Came Back

Anthropic published lab-validated results on 18 August from a protein design campaign it ran with minimal human involvement. Claude was asked to design 30 minibinders for each of 15 protein targets. A minibinder is a small protein engineered to latch tightly onto a specific target protein, which is how a great many modern medicines actually work. Two independent labs, Adaptyv Bio and Twist Bioscience, physically made and tested the designs. Of 1,320 designs, 354 bound, and they covered 14 of the 15 targets.

The hit rate is the number to look at. Anthropic reports 22.6 percent for Opus 4.8 and 26.7 percent for Mythos Preview when designing against all targets at once inside a single 48-hour session, rising to 35.1 percent for Mythos Preview when each target got its own 24-hour session. Anthropic’s stated baseline for protein design campaigns today is 10 to 15 percent. Against RBX1, a target Adaptyv Bio has run a public competition on, Mythos Preview hit 40 percent where human entrants averaged 3.7 percent, and its best design outperformed the competition winner. Against TNFα, the target behind drugs like Humira, Opus 4.8 produced binders that worked across human, monkey and mouse versions of the protein, which matters for animal studies. Curiously Mythos Preview failed on that one and Anthropic says it does not know why.

Here is what is genuinely new, and it is not that an AI designed a protein. Specialist machine learning models have been designing proteins for years. What Claude did was operate those specialist models: choosing where on each target to aim, generating candidate structures, running optimisation cycles, and screening for candidates that would actually express and stay soluble. That orchestration work is normally weeks of a computational biologist’s time. Claude got a 30,000-token prompt, internet access, GPUs and a fixed time budget, and then ran unattended. The humans approved network requests and ordered the physical tests.

An honest caveat list, because Anthropic supplies one. Claude failed completely against maltose binding protein, where none of 90 designs bound, and managed only weak binders against BBF-14. A high-affinity binder is the first step in drug development, not the last, and Anthropic says the results still need further characterisation. There is also a real dual-use problem, which the company addresses directly: the same capability that speeds up medicine speeds up bioweapon research, which is why protein design stays blocked in its most capable model while it builds a vetted access programme for scientists.

What this means for you: for most people, nothing changes today, and no drug exists because of this. What it tells you is where the useful shape of AI in science is landing. The win was not a smarter model replacing a scientist, it was a general model driving a stack of narrow expert tools competently for two days straight, a job that previously required a specialist’s full attention. Anthropic’s second result in the same post makes the same point more modestly: given raw instrument files and a two-sentence prompt, Claude Opus 5 processed chemistry lab data in 23 minutes, landing on 96.4 percent purity against the lab’s own 96.33. That is not a breakthrough. It is a week of somebody’s tedium removed, and there is rather a lot of that to go round.

Sources

• Anthropic: How Claude is accelerating protein design and analytical chemistry
• Adaptyv Bio: Benchmarking Claude’s protein designs in the wet lab
• Anthropic: protein design technical report