On August 18, 2026, Anthropic published an official research post describing how Claude performed on two experimental tasks tied to the early stages of drug discovery. The first involved designing brand-new protein binders from scratch against specific targets; the second involved analyzing the structure and purity of chemical compounds. Both experiments were validated against real laboratory data.
How did the protein binder design experiment work?
Claude Opus 4.8 and Mythos Preview (a research-only preview build of Anthropic's Mythos model line) designed new protein binders against 15 target proteins, succeeding against 14 of them. A protein binder — a small protein designed to latch tightly onto a target protein — is central to how many drugs work, since they attach to a target to inhibit, activate, or deliver something to it. This was de novo design — designing something new from scratch, without an existing structure to work from — a task that has historically taken protein engineers months of computation, optimization, and screening per target.
Claude's designs were independently produced and tested by external evaluators Adaptyv Bio and Twist Bioscience in a wet lab — a physical laboratory where scientists handle real chemical and biological reactions with test tubes and cell cultures, as opposed to a computer simulation. At least six targets yielded high-affinity binders (affinity is a measure of how strongly a protein binds to its target), and at least four targets produced binders matching or exceeding the best previously reported affinity. The post notes some designs bound several times more tightly than the best published result.
Caution: the official post is explicit that while the design step got faster, validating that data in a wet lab still takes weeks. This experiment demonstrates speed at the design stage, not a shortened drug-development timeline overall.
How did the success rate compare with the industry norm?
When designing against all 15 targets simultaneously in a 48-hour session, Mythos Preview achieved a hit rate — the share of designed candidates that actually bound their target — of 26.7%, and Opus 4.8 achieved 22.6%. According to the post, hit rates of 10 to 15% are typical in protein design campaigns today. Separately, the post's summary states that, depending on the setup, between 22% and 35% of individual designs bound successfully.
| Group | Hit rate |
|---|---|
| Typical industry campaign | 10-15% |
| Claude Opus 4.8 (48h, 15 targets at once) | 22.6% |
| Claude Mythos Preview (48h, 15 targets at once) | 26.7% |
What happened in the chemistry analysis experiment?
Claude Opus 5 — a generally available model anyone can use today — was given only a contract lab's raw data files and a two-sentence prompt, and finished analyzing NMR (nuclear magnetic resonance spectroscopy — a technique for determining how atoms are arranged inside a molecule) and LC-MS (liquid chromatography–mass spectrometry — a technique for separating and measuring a compound's components and purity) data in 23 and 19 minutes respectively. The results matched the lab's own analysis on both hydrogen counts and purity (96.4% versus 96.33%).
Unlike the protein design experiment, which used the not-yet-generally-available Mythos Preview, the chemistry analysis experiment used Claude Opus 5, which is currently accessible on Claude.ai and via the API. This example demonstrates what a generally available model can already do, not a research-only capability.
What do these results actually mean?
The official post frames these results as speeding up parts of the early stages of drug development, while noting that accelerating drug development end-to-end depends more on policy and operational bottlenecks than on further gains in core scientific capability. It also states that life science research tasks are currently blocked in Anthropic's most capable model, and that launching an access program for scientists is a high priority, with more to be shared soon. The most capable model generally available today is Opus 5.
As background, the same post notes that in mathematics, decades-old Erdős problems are now being solved at a rate of several per month, and that Claude recently improved a longstanding lower bound on the Riemann zeta function. The post frames this protein and chemistry work as part of a broader trend of AI contributing not just to fields where verification is fast, like math, but increasingly to experimental fields like the life sciences, where verification is slower and more expensive.
Frequently asked questions
Q. Can I run an experiment like this myself with Claude Opus 5 today?
No. According to the official post, life science research tasks are currently blocked in Anthropic's most capable model, and the protein design experiment used Mythos Preview, which is not yet generally available. Anthropic says it is preparing a separate access program for scientists. Tasks like the chemistry analysis, however, are already achievable with Opus 5, which is generally accessible.
Q. What's the difference between hit rate and affinity?
Hit rate is the share of designed candidates that actually bound their target, while affinity measures how strongly a bound pair sticks together. A high hit rate doesn't guarantee high affinity, and vice versa — they are separate metrics.
Q. Does this mean a new drug is coming soon?
No. The official post describes this as speeding up specific early-stage tasks in drug discovery, while noting that wet lab validation still takes weeks and that policy and operational bottlenecks play a larger role in the overall drug development timeline.