I am not a biochemist. Show me a chart of binding affinities and the most I can tell you is which bar is taller. Still, I kept coming back to the research Anthropic published on Tuesday, and not because of the proteins. What holds my attention is the shape of the experiment: an agent got an assignment, decided for itself which specialist models to install and run, and the output was then physically rebuilt and measured by two outside labs.
That last part is the point. Almost everything I write about agents, and everything I build with agents myself, gets graded by the same party that asked for it. Here the grading was done by lab equipment that has no idea a language model was involved.
What actually happened
On 18 August 2026 Anthropic published two results in one post. The first is about protein design, the second about analytical chemistry.
For the protein work they picked 16 targets, including all of Adaptyv Bio’s BenchBB set plus two targets from recent design competitions that can barely be in training data yet. Claude was asked to design 30 minibinders per target: small proteins that latch onto a larger one. That latching is how a large share of medicines work in the first place.
The designs went to Adaptyv Bio and Twist Bioscience, who produced and tested them independently of Anthropic. One of the 16 targets produced no usable measurement because the target itself aggregated, which leaves 15. Claude found working binders against 14 of them.
The models doing the work were Opus 4.8 and Mythos Preview, running inside Claude Science. The chemistry half ran on Opus 5, the model anyone can use today.
The mechanism: the agent operates the tools, it does not replace them
This is the part that drops out of the summaries. Claude did not conjure proteins out of nothing. It drove the existing, openly available design and co-folding models the field already relies on. Reading up on each target, choosing where on the protein to bind, installing and running the specialist models, filtering candidates, improving them, and finally picking the thirty that go to the lab.
That is exactly the layer that used to require a computational specialist. Not the heavy math itself, but the orchestration around it: which tool, in what order, at what threshold, and when to abandon a branch.
The setup, in numbers:
| Multi-target | Single-target | |
|---|---|---|
| Models | Opus 4.8 and Mythos Preview | Mythos Preview |
| Approach | all targets in one session | one session per target, in parallel |
| Wall time | 48 hours | 24 hours per target |
| Compute budget | up to 12,500 H100 hours | up to 2,500 H100 hours per target |
Human involvement, per Anthropic, came down to approving requests Claude made (network access, code execution), fixing infrastructure problems outside the design sessions, and ordering the designs from the labs. The substantive prompt was human work: roughly 30,000 tokens of protocol written by an expert, published openly on Hugging Face along with the data.
The numbers, including the awkward ones
- 354 working binders out of 1,320 designs with usable measurements. That is 26.8 percent, and I did that division myself. Anthropic quotes a range of 22 to 35 percent depending on the setup.
- The field typically sits at 10 to 15 percent. That figure comes from Anthropic’s own derivation off proteinbase.com, not from an independent meta-analysis.
- Per target the hit rate runs from 90 percent down to zero. The average hides an enormous spread.
- Against RBX1, Mythos Preview hit 40 percent in single-target mode, against 3.7 percent among entrants in Adaptyv Bio’s competition for that target. Its best design beat the winning entry, which came out of a field of 245.
- Against MBP: zero out of 90. Maltose binding protein is large, flexible and smooth, so there is very little to grab. One design showed a weak but reproducible signal, and that was it.
- Against BBF-14, three binders with modest affinity. That one is a β-barrel that was itself computationally designed, which is precisely why it sits in the benchmark.
- TNFα only worked with Opus 4.8, not with Mythos Preview, even though Mythos Preview is the stronger model overall. Anthropic states plainly that they do not know why.
Those last two lines interest me more than the 26.8 percent. A system that succeeds on a hard target with the smaller model, with nobody able to explain the reason, is not a system you build a project plan around.
Rohan Paul, who worked through the technical report, cites a split I cannot find in the blog post: 35.1 percent in single-target mode against 26.7 percent when targets shared a session. That fits inside Anthropic’s stated range, but I could not read the report myself (the PDF would not parse), so I carry it as his reading rather than a verified fact. Same goes for his observation that the system cannot reliably tell when a campaign has failed: targets that produced nothing scored much like targets that worked.
The second result, which got less attention
Alongside the protein work, Opus 5 was handed a contract lab’s raw measurement files, an NMR spectrum and an LC-MS run, with a two-sentence prompt attached. No explanation of the file format, because that format belongs to the instrument vendor and is undocumented.
Claude worked out how the data was encoded, checked itself by reproducing the instrument’s own recorded totals across all 2,664 scans, and then delivered what a chemist would deliver: the separation trace, the spectra, a purity table, the molecular mass, plus reusable code for reading files like these. Purity 96.4 percent against the lab’s own 96.33 percent. Turnaround of 23 and 19 minutes, where the lab’s finished report arrived four days after the first spectrum.
For my own work this is the more relevant result. The protein work ran on models nobody can request. This ran on the model sitting in my terminal.
The caveat
The standard shifts from “can an expert do this” to “can an expert check this”. With proteins the check is still honest, because a wet lab is a brutal test: it binds or it does not. But Anthropic says outright that it intends to characterise the hit rates and affinities more extensively. So publication ran ahead of the final measurement, which is the same pattern I keep seeing in my own field. Generation gets faster, validation does not, and in that gap people start trusting results nobody has checked.
Responsibility lands with the party that can see the least. Anthropic names the dual-use problem and keeps protein design outside general access in Fable 5, routing it through access programmes instead. That is a serious choice, not a marketing line. It also means the body deciding who may use this is the vendor, not a regulator and not the field. In public sector innovation that tension is familiar: the gatekeeper and the builder are the same organisation.
Judgement drains out of the intermediate steps. A computational biologist running their own pipeline notices where it goes sideways along the way. When the agent produces 1,320 designs and selects thirty per target on its own, that running commentary disappears. The finding that the system does not recognise its own failed campaigns makes it concrete. Nobody steps in halfway through any more, and the system gives no signal that someone should.
What I do with it
Nothing involving proteins. Plenty involving the framing. I am treating this as the sharpest example yet of something I run into constantly: an agent that operates other systems is a different promise from an agent that answers questions, and the quality lives in the orchestration rather than in the final model.
Where I apply it directly is verification. In my own agent setups I have a second agent try to wreck what the first one built, and this research hands me an argument I did not have: the system is worse at judging its own failures than at producing successes. That is not something a better prompt fixes. You fix it by putting the test outside the system, the way Anthropic did here with two external labs.
The chemistry half goes straight into my practice: raw measurement files in an undocumented format, with instructions to reproduce the source’s own totals before interpreting anything. That check up front decides whether you can trust the rest of the answer.
Frequently asked questions
Does this mean AI-designed drugs are on the way?
No, and Anthropic makes that point most clearly of anyone. A binder is not a drug. Designing a tightly binding molecule is the first step, and designing a drug is itself one phase out of the many needed to show that something is safe and effective. Minibinders are also not a standard therapeutic format.
Can I try this myself?
Not the protein half. Those capabilities are deliberately outside general access and run through access programmes for researchers. The prompt and the data are published openly on Hugging Face, so you can read the setup. The chemistry half ran on Opus 5, which is generally available.
Why is 14 out of 15 more impressive than it sounds?
Because the targets were chosen to be difficult, including two that can hardly appear in training data, and because the test happened outside the model. This is not a benchmark score the model computes about itself. These are molecules two external parties built and measured.
Where is the biggest uncertainty in these numbers?
In the spread and in the baseline. Per-target hit rates run from 90 percent to zero, so an average tells you little about the next target. And the 10 to 15 percent reference comes from Anthropic’s own derivation rather than an independent source.
Sources
- “How Claude is accelerating protein design and analytical chemistry”, Anthropic — 18 August 2026: anthropic.com
- Anthropic on X, thread with the announcement, the hit rates and the note that binders are not drugs — 19 August 2026: x.com/AnthropicAI
- Rohan Paul on X, summary of the technical report including the per-mode split — 19 August 2026: x.com/rohanpaul_ai
- Prompts and data from the design campaign, published openly — accessed 19 August 2026: huggingface.co
- Protein design technical report (PDF) — accessed 19 August 2026: www-cdn.anthropic.com
Fact-checked on 19 August 2026. Every figure on targets, hit rates, compute budget and the chemistry analysis comes from Anthropic’s own post and was checked there rather than taken from coverage of it. The 26.8 percent is my own division of 354 by 1,320, a number Anthropic does not state in that form. Two things I could not verify. The technical report is a PDF I could not parse, so the 35.1 against 26.7 percent per-mode split and the observation that the system fails to recognise its own failures rest on Rohan Paul’s reading, not mine. And the “10 to 15 percent is typical” comparison leans on a derivation Anthropic made from proteinbase.com; I found no independent source confirming that range. The measurements were carried out by Adaptyv Bio and Twist Bioscience, but the reporting on them comes from the company that made the model.