On September 23 Anthropic said Claude had found something new in biology . Roughly 950 agents spent 21 hours and 210 million tokens hunting through genetic databases, working down from more than 200,000 known genes to 3,500 arrangements worth a second look, and then to the 20 most promising. One of the 20 was a layout no published study had described: DNA repeats spaced at regular intervals, with genes sitting alongside. Claude measured the spacing, compared it against known systems and searched the literature for any earlier report. The layout resembled the architecture CRISPR is built on, which is why it caught their eye.

CRISPR is gene-editing technology: it finds one exact spot in a genome of three billion letters and changes what is written there. Control-F and Control-V, for DNA.

Gene-editing stocks fell the same day . Prime Medicine dropped 11.58%, Beam Therapeutics 6.02% and CRISPR Therapeutics 5.54%. Biotech was having a bad day regardless, with the 10-year Treasury yield pushing above 5.11% that afternoon, its highest since 2007, and selling running across the sector. So the prices can’t tell us how much of that was Claude, but the headlines made the connection anyway.

That reflex is becoming routine. In February, when Anthropic released a model it said could produce the kind of financial analysis that normally takes people days, FactSet fell as much as 10% on the day, S&P Global 4.7% and Moody’s 2.9%. In July, when OpenAI launched Presence, its agent product for corporate workflows, the IGV software index dropped 3% before the day was out. Each time a sector was marked down within hours of a company blog post.

The difference this week is that there was nothing to buy. In February and July there was a product on the other end of the selloff, and on Wednesday there was a finding whose function Anthropic can’t really name — it can’t say what the system Claude found actually does, and says the work to find out is ongoing.

Anthropic’s description of who did what is one of the most interesting sentences in the announcement: “Our involvement was limited to the initial prompt and the lab work, while Claude agents combed through the database.”

Everything in between the prompt and the lab work was the model, and everything that touched a bench was human.

Five days earlier Anthropic had confirmed something it had kept quiet: it runs a wet biology lab in the Bay Area. Its head of life sciences, Eric Kauderer-Abrams, told Reuters that “to do biology, the final test is still and will be for a while in real lab work.”

Then the disclosure stopped: Anthropic won’t say how big the lab is, how many people work there, or when it opened. We know the exact token count because Anthropic wanted us to know it, and we don’t know whether the lab is one bench in a corner or an entire floor. This can be read two ways, a pilot too small to announce or a bet too early to describe, but it’s impossible to say which.

Where The Hypotheses Pile Up

The search took 21 hours. Choosing which of the 20 candidates deserved an experiment, and then running those experiments, took people. The company describes the loop it built around that: Claude produces hundreds to thousands of hypotheses at a time, its scientists judge which are worth testing, and that judgment is written back into Claude’s instructions so the next batch comes back closer to what they would have picked themselves.

Coming up with a plausible idea used to take somebody half a career, and now it costs tokens. But testing one costs what it did before. A pipette takes as long as a pipette takes, and somebody still has to be standing there.

I wrote here two weeks ago about OpenAI putting ten thousand agents on the Navier-Stokes problem for 88 hours and getting back a 166-page proof a machine could check and a person could not yet understand. The agents were the cheap part in that story too, and what came after them was somebody reading.

Pharma has been living this for decades. Roughly 23,000 drugs sit in development worldwide , and the FDA approves about 50 new ones a year. Measure that against CRISPR itself: 11 years separated the paper that showed what CRISPR did from an approved medicine built on it, and this find has not reached the point where that clock would start.

A Model Company Bought A Bench

The lab is a bigger story than anything Claude found in that database. Anthropic sells a model; It has no manufacturing, no physical product, no reason to own a room full of centrifuges, and it built one anyway. Dario Amodei says Claude might eventually run the equipment itself, “but we aren’t doing that today.”

Owning a lab doesn’t mean doing everything in it. Anthropic still sends biological work to outside partners, and its own lab runs at the lower biosafety levels, so some experiments can’t happen there at all. What it changes is which ones need somebody else’s permission. An idea you can put on your own bench this week is a different kind of idea from one that needs a purchase order and a slot in a vendor’s calendar, and Claude is turning them out by the thousand.

Look at what these companies have been buying: In April, Anthropic committed more than $100 billion to Amazon Web Services over the next decade and secured up to five gigawatts of capacity, on top of the more than 1 million Amazon Trainium chips it already uses to train and run Claude. OpenAI designs its own accelerators now, under a deal with Broadcom to deploy 10 gigawatts of them starting this year. In January, Meta signed nuclear agreements for more than 6 gigawatts, including the output of the Davis-Besse and Perry reactors in Ohio.

Chips, buildings and electricity all scale with the size of the check. A bench does not, which is exactly where Anthropic decided to spend.

What This Means If You’re Funding Any Of This

I look at founders pitching “the AI layer for biology,” and the deck is almost always about model architecture and proprietary training sets. That pitch has a hole in it now, as the frontier model builders themselves just signaled that intelligence alone is insufficient.

In this new regime, software generates an infinite queue of hypotheses for the price of electricity. The actual bottleneck instantly shifts to physical throughput.

The question to now ask first is unglamorous: how quickly can you turn a hypothesis into a measured result, and how much of that loop do you control? Access to frontier models is increasingly becoming a commodity with a price list, while bench time isn’t. A company whose hypotheses sit in a queue waiting for someone else’s lab schedule and whose failures produce no proprietary learning, has a different business from the one in the deck.

It’s difficult to say how long this holds. If self-driving labs get good quickly the constraint dissolves, and the advantage goes back to whoever generates better hypotheses — which is plainly the direction Amodei is betting. My bet is the other way: the company with the most information about what these models can do just spent money on the assumption that humans will be holding the pipettes for a while.