Anthropic Built a Wet Lab. The Bottleneck in AI Drug Discovery Was Never the Model.

On September 18, Anthropic confirmed it is running a wet biology lab in the San Francisco Bay Area where Claude models help design and run physical experiments. The same day, the South China Morning Post reported that Anew Labs, a ByteDance spin-off working on AI drug design, raised $290 million at a $1.5 billion valuation. Three days later, HHS announced ARPA-H investments in AI and computational modelling as part of a shift away from animal testing.

Three stories, one theme. In AI biology the model was never the scarce part. Checking what the model says is.

## Predictions are cheap, confirmation is not

Reuters' framing of the Anthropic lab is the useful one: it exists to test AI-generated hypotheses and close the loop between model predictions and real-world biology. Anthropic stresses the facility does preclinical research, not direct drug development, and pairs it with a Life Sciences Verification Program. The word in that programme name is the point.

Compare software. Anthropic says Claude now leads 26 percent of its own AI research. That loop closes fast because a model can check its own code in seconds. In biology the check is a bench, reagents and weeks of waiting. A lab that can design an experiment in minutes still has to run it. So the company bought pipettes.

## The money agrees

Anew Labs is the largest of a run of rounds. ByteDance keeps a 56 percent stake, and the investors are HSG, IDG Capital and 5Y Capital. Earlier this month Superluminal Medicines raised a $60 million Series B led by BVF Partners to push a GPCR obesity programme into a Phase 1 trial. And Beijing's national AI medical pilot base launched a 100 million yuan biopharma plan that comes with more than 1 billion yuan worth of compute for drug discovery and clinical development.

Notice what each one pays for. Superluminal's round funds a trial, the most expensive form of verification there is. Beijing's subsidy is mostly compute, which is the cheap half. Anew's $290 million, per the SCMP report, funds a platform, and the question for its investors is how much of that becomes bench capacity.

## Washington moves the goalposts

The HHS package matters because it changes what a validated prediction is worth. FDA, NIH and CDC are updating rules and funding to promote non-animal methods supported by AI, with ARPA-H funding computational models under its CATALYST programme. If a regulator accepts a modelled result where it once required an animal study, the output of a well-verified model becomes a regulated asset rather than a research shortcut.

Two hedges. The announcement is a package of actions, not an approved pathway, and our record does not state a dollar figure for CATALYST. And the same capability cuts the other way: Anthropic's own September threat report described blocked attempts to use Claude for biological weapons research. A lab that makes Claude better at biology raises the bar on its safeguards at the same time.

## What to do with this

First, when you look at an AI drug company, ask for its verification throughput before its benchmark scores: how many hypotheses it can physically test per week, and who runs the bench. That number, not the model, sets its speed. The rounds in our investment tracker rarely state it, which is itself informative.

Second, watch for the first FDA acceptance of an AI-supported non-animal package under the HHS shift. That date, whenever it comes, is when a verified prediction starts to carry regulatory value, and it will reprice every company on the Claude biology trend we are tracking.