Anthropic has confirmed it operates a wet biology lab in the Bay Area where its models run physical experiments. The disclosure came a day after the Life Sciences Verification Program launched, relaxing Claude's safeguards for vetted biologists. A company warning about biological AI risk is now a user of it.
Key takeaways
- Anthropic confirmed a Bay Area wet lab where AI models run physical experiments
- The lab focuses on fundamental biology, not drug discovery
- April 2026: Coefficient Bio acquired in an all-stock deal reported at about $400 million
- 17 September 2026: Life Sciences Verification Program launched with Mythos, Opus and Sonnet access
- 16 September 2026: drug discovery deal with Novo Nordisk
The physical experiment as the final test
In his September essay "We Must Pace the Frontier", Dario Amodei wrote that AI could cure most major diseases within 5–10 years. The condition is a way to test hypotheses in the real world, and an in-house lab answers it.
We believe that to do biology, the final test is still, and will be for a while, in real lab work. We absolutely are doing that today.
Eric Kauderer-Abrams, head of life sciences at Anthropic, quoted by Reuters.
Verification instead of blocking
The program splits access into two tiers. Standard use is annual team-wide access. High-risk use covers a single project, renews every six months, and for Mythos-class models also requires US government vetting. Anthropic uses offline monitoring?offline monitoring: Oversight based on reviewing usage records after the fact, rather than blocking prompts at the moment they are sent., not real-time blocking. A same-day report shows the payoff: Claude optimised over 30 open-source biomolecular models?biomolecular models: Computational programs that predict the structure and interactions of biological molecules such as proteins and nucleic acids. in under four weeks, roughly fourfold faster on average.
| Tier | Scope | Renewal | Extra condition |
|---|---|---|---|
| Standard use | Whole team, routine research work | Annual | None |
| High-risk use | Single project | Every six months | US government vetting for Mythos-class models |
The inconsistency charge
Chamath Palihapitiya wrote on X that "the group behind such hits as 'We're All Going To Die' and 'Regulate Me Now' are building a wet lab in SF". Nine days earlier, departing researcher Jacob Coxon said those building AI earnestly believe it could kill everyone by the end of the decade.
Why it matters
Anthropic stops being only a model vendor and becomes an entity generating its own wet data?wet data: Measurements produced by physical laboratory experiments, as opposed to data generated purely computationally.. That shifts the axis of competition: the edge comes not from model quality but from a closed hypothesis-experiment-result loop rivals lack. The company also configures for itself the safeguards it promotes as an industry standard, without external oversight. For regulators this is a new kind of actor, model vendor and laboratory at once.
What next?
- Individual access on Pro and Max plans follows the teams and institutions phase
- Anthropic expects hundreds of verified organisations in the program's first week
- High-risk permissions require renewal every six months
Sources
- TechCrunch — Anthropic is operating a lab that conducts biology experiments
- Anthropic — Introducing the Life Sciences Verification Program
- Anthropic — How Claude is uplifting biomolecular modeling
- Novo Nordisk — Novo and Anthropic will collaborate to advance drug discovery with Claude
- TechCrunch — Anthropic buys biotech startup Coefficient Bio in $400M deal
