Anthropic announced last Wednesday that a molecular biology lab it quietly launched earlier this year — where Claude agents read and reason about hard biology problems while human scientists run the experiments the agents propose — has made its first discovery. The claim has not survived contact with the biologists best positioned to evaluate it.

What Anthropic says its system did: 950 Claude agents spent 21 hours combing a library of millions of DNA sequences and flagged a repeating pattern surrounding a known enzyme. Anthropic itself acknowledges the agents did not find a brand-new sequence. The company says the particular pattern had not been catalogued before — but its announcement leans hard on the framing, calling the find "reminiscent" of what led to CRISPR, the gene-editing technology that "has already transformed science and medicine."

For context on the actual workflow: the scenario Anthropic describes is one where a researcher flips through massive sequence libraries, hunts for a peculiar sequence encoding an interesting enzyme, then works out what that enzyme does and how to manipulate it. The agents handled the first, pattern-spotting step. The hard downstream biology — determining mechanism and function — remains untouched.

That distinction drew a viral critique post from Lucas Harrington, a biologist, later endorsed by the chair and CEO of the drugmaker Eli Lilly. "Finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does," Harrington wrote. In his framing, the agents performed laboratory grunt work. A discovery it is not.

The situation then got messier. Mario Rodríguez Mestre, a biologist at the University of Copenhagen, told the New York Times over the weekend that his team had already discovered this particular pattern. Mestre, who regularly chatted with Claude in his own work, says he now wonders whether Anthropic's team learned from those conversations. Anthropic denies it. Mestre says he is stopping all use of Claude regardless — a data-provenance question that any lab feeding proprietary or unpublished results into the model should weigh directly.

The episode turns on a framing choice by AI vendors. These companies are not presenting their systems as instruments, like microscopes or supercomputers, but as agents making discoveries themselves — a framing many researchers find incompatible with how scientific knowledge actually accumulates, through collaboration and an expanding toolset.

There is a real result buried under the marketing, and R&D managers evaluating agent tooling should not dismiss it. Whittling 200,000 candidates down to a few worth pursuing is legitimate scientific work, and the fact that a general-purpose model did it — with human steering and humans running the experiments — is notable. Narrowing 200,000 candidates to a shortlist has direct implications for screening budgets and triage time in any sequence-heavy program. But by setting the bar at "AI made a discovery," Anthropic converted a credible triage result into ammunition for a binary argument: breakthrough or bust.

The same dynamic is playing out in mathematics. Earlier this month OpenAI said its team agents had cracked a million-dollar problem. Weeks later, skeptics circulated a Scientific American article asking whether OpenAI had solved the wrong Navier-Stokes problem. That piece did not dispute the solution's correctness; it argued the specific result may not be the one mathematicians care most about. Add a mathematician's accusation that the models used some of his work without credit, and observers split into camps: OpenAI cheated, or the result didn't matter — or both.

Harrington closed his critique with a suggestion aimed squarely at vendor incentives: AI companies should "set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is." With Sam Altman and Dario Amodei racing to one-up each other on scientific-claim headlines, that higher bar is unlikely to arrive on its own. Labs adopting these agents would be well served by treating vendor "discovery" claims the way Mestre and Harrington did — as inputs to verify, not results to bank.