News · 2026-09-01

Claude designed protein binders that worked about half the time

Anthropic gave Claude Mythos 5.1 access to open-source protein design and folding tools, told it to design molecules that stick to specific biological targets, and sent the results to two outside organizations for laboratory testing. Nearly half of the designs bound successfully across 12 targets. The normal hit rate in protein design is 10 to 15 percent. On three of those targets, the model's designs bound ten times more tightly than the best entries submitted to Adaptyv Bio's public protein design competitions.

Key facts

• Roughly 50 percent hit rate across 12 targets, against a field-typical 10 to 15 percent, with all designs experimentally validated by outside labs.
• Ten times higher binding affinity than the best submissions to Adaptyv Bio's design competitions, on three targets.
• Published September 1, 2026 by Anthropic, alongside the Fable 5.1 and Mythos 5.1 release.
• Primary source: Anthropic's launch post.

Protein binders are the working end of a large fraction of modern medicine. A drug that blocks a receptor, activates a pathway, or delivers a payload usually does it by physically gripping a target molecule, and the tighter the grip, the smaller the dose you need. Designing one from scratch is mostly a numbers game: you generate many candidates, most of them do not stick, and you find out which ones work only by making them and testing them in a lab. That failure rate is the cost center. Cutting it from roughly nine misses in ten to roughly one in two does not just make the work faster -- it changes which projects are affordable at all.

The mechanism here is worth being precise about, because the easy misreading is that a language model invented new biology. It did not. Anthropic handed the model the same open-source design and folding tools human researchers already use, and the model drove them: choosing what to try, reading the results, and iterating. What is new is the loop, not the chemistry. The closest analogy is the difference between owning a well-equipped workshop and having someone in it who never gets tired of trying the next thing. Our explainer on de novo protein design covers what those underlying tools actually do.

The same post describes two other results in the same shape. Claude Fable 5.1 trained a neural network on radar images that NASA's Magellan mission collected more than 30 years ago, plus an existing elevation map covering a fifth of the planet, and produced a new elevation map of roughly a third of Venus. The old map resolved features at 10 to 20 kilometers; the new one resolves them at two to three, with heights up to 25 percent more accurate. Anthropic released it under a Creative Commons license, timed ahead of NASA's VERITAS and ESA's EnVision missions, in the hope that mission planners use it to pick observation targets. That is data that has been sitting in a public archive for three decades.

The third result is the least glamorous and possibly the most useful. Computational biologists run specialized machine learning models on GPUs constantly -- testing, for instance, every possible mutation near every human gene means running the same model thousands of times. Mythos 5.1 wrote custom GPU kernels and added caching for seven open-source models, including ChromBPNet, Enformer, ProGen2, and both the 7-billion and 40-billion parameter versions of Evo 2, speeding them up by as much as 2.5 times with identical outputs. On genome-wide analyses, Anthropic estimates that cut GPU costs by 30 to 60 percent -- one analysis dropping from about $30,000 to $21,000, another from $18,000 to $8,000. "This kind of optimization would normally take a team of performance engineers weeks, and is often unaffordable for academic labs," the post says. The model did it in days from public source code. Anthropic says it plans to open-source the optimizations.

Why this matters beyond the headline: all three results are examples of the same thing, which is an agent running a long, unglamorous research loop that a human would find tedious and a grant would find hard to fund. None of them required a scientific insight the model invented. All of them required patience with existing tools and existing data. That is a narrower claim than "AI does science," and a more credible one. It also lines up with why Anthropic opened a hardware standard for Claude to operate lab equipment last week -- the loop only closes if the model can run the experiment too.

The honest caveat is a large one. Every number here comes from Anthropic's own announcement about Anthropic's own model, published on launch day. The protein binder work was validated by external labs, which is the strongest evidence in the set, but the organizations are not named and no paper or preprint accompanies the claims. The Venus map is public and checkable; the GPU kernels are promised but not yet released. Until the optimizations ship and someone outside the company reproduces the binder hit rate, this is a well-specified claim rather than a confirmed result. It is also worth noting that the more permissive Mythos 5.1 -- the version that did the biology work -- is available only to vetted organizations, so most researchers cannot try it themselves.

Primary source, verified: read the paper →

Ground Truth. (2026, September 1). Claude designed protein binders that worked about half the time. Ground Truth. https://groundtruth.day/news/claude-designed-protein-binders-that-worked-half-the-time.html

@misc{groundtruth:claude-designed-protein-binders-that-worked-half-the-time,
  title  = {Claude designed protein binders that worked about half the time},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/claude-designed-protein-binders-that-worked-half-the-time.html}
}

Topics: anthropic · claude · ai-for-science · protein-design · biology · astronomy · gpu