Anthropic puts its biology claims on the bench in a new Bay Area wet lab

Somebody at a biotech has to decide whether a model's suggestion is worth the reagents and the week. One line in the Reuters report speaks directly to that decision. "We believe that to do biology, the final test is still and will be for a while in real lab work," Eric Kauderer-Abrams, Anthropic's Head of Life Sciences, said ^[3]. He added: "We absolutely are doing that today..." ^[4]

Kauderer-Abrams did not specify what the lab works on ^[8], and a spokesperson told Reuters the lab was not created specifically, or perhaps only, for drug discovery ^[7]. Anthropic launched a drug discovery program called Claude Science in June, saying it wanted to pursue research pharma companies would not touch because it likely would not pay off financially ^[5][6]. Kauderer-Abrams said he believes AI could speed the discovery of treatments for conditions previously thought too difficult to treat, including complex and targeted therapies ^[13]. Nobody outside can inspect the bench work behind that yet.

Kauderer-Abrams also said Anthropic does lab work in-house while working with external partners when it makes sense ^[9]. Anthropic is running both tracks, and it has not said how the work divides between them. The whole account rests on a single Reuters report, relayed by Engadget, with no second publisher confirming it independently ^[16].

Sources told Reuters that one thing Anthropic is looking at is having Claude control robots to carry out experiments ^[10]. Kauderer-Abrams placed that work "in the very early innings of using AI to automate the execution of lab work" ^[11] and said human oversight remains essential for safety, at least for now ^[12]. He also said Anthropic has yet to run any clinical trial ^[14]. For a customer, those are the two dates to ask about: the first robot-executed experiment series, and the first trial.

The lab arrives with Anthropic's own bio disclosures behind it. The company recently said it flagged and stopped multiple scientists who were using its models to create potential biological weapons ^[15], and chief executive Dario Amodei has called for a slowdown in advanced AI development to reduce the risk of losing control of frontier systems ^[17]. A company that says both of those things while standing up bench capacity should expect to be asked who inside it can authorise an experiment.

Two axes sort any AI-for-science claim you are being sold. First, was it tested in a physical lab or on a benchmark. Second, does the party that ran the test stand to gain from the result. Anthropic's own statement moves its biology work onto the physical side of the first axis and leaves it on the interested side of the second. The contractual answer to that is a replication clause: a named outside lab repeats the key experiment, at the vendor's expense, before the result goes in a slide.