What did Anthropic actually build, and why does a wet lab matter?
Anthropic opened a physical wet lab in the Bay Area this month so Claude can move past talking about biology experiments and start helping run them directly, with real equipment and real results.
A wet lab is a real room with real equipment: pipettes, incubators, sequencing machines, the physical gear scientists use to run experiments on living cells and molecules.
Until now, an AI model's contribution to biology mostly stopped at the screen. It could read papers, suggest a hypothesis, or predict how a protein might fold. A human then took that prediction into a lab and found out if reality agreed. That handoff is where a lot of promising computational biology quietly dies. A wet lab closes the gap, letting Claude's predictions get tested faster by the same team that made them.
This is not just branding. Anthropic already has numbers to back up the claim that Claude's contribution goes beyond chatting. Working inside Claude Science, the model optimized more than 30 deep learning models used for biological tasks in under four weeks. On average, those models ran roughly 4x faster afterward, with only a small accuracy trade-off, and nearly 2x faster with no accuracy loss at all.
How did Claude speed up biology software that fast?
Claude also built something called a low-memory mode, nicknamed "Big," that lets a single graphics processing unit handle biomolecular structures far larger than before.
A graphics processing unit, or GPU, is the kind of computer chip built to crunch huge amounts of data at once. Folding a large protein or predicting how a huge molecular complex holds together normally needs several of these chips working together, because the math gets too heavy for one to handle. Claude's low-memory mode changed that math, letting researchers accurately model systems bigger than 10,000 amino acids and nucleotides, and run predictions on systems past 70,000, using a single GPU node instead of a cluster.
The practical result: researchers used it to fold structures like human mitochondrial complex I and a bacterial ribosome, both large, complicated molecular machines, on hardware that used to be too small for the job. Fewer chips needed means more labs can afford to run this kind of work, not just the ones with the biggest budgets.
| Step | Before | What's changing |
|---|---|---|
| Hypothesis | Claude suggests an idea from literature | Same, but faster iteration with lab feedback |
| Prediction | Model predicts structure or outcome | Runs on cheaper, smaller hardware via low-memory mode |
| Validation | Separate outside lab tests the prediction, on its own timeline | Anthropic's own wet lab can test it directly |
| Access | Fable-class models block biology and drug-development queries by default | Life Sciences Verification Program grants vetted researchers deeper access |
If AI is this fast at biology, why haven't we seen AI-discovered drugs yet?
Speed was never the bottleneck holding back AI-designed drugs. Proof is, and proof still requires testing a compound inside an actual living patient over time.
More than $40 billion in venture capital has gone into AI-native biopharma companies this decade, according to PitchBook data. Despite that money, almost no AI-associated drug candidate has made it past Phase 2 clinical trials, and as of mid-2026 no AI-discovered or AI-designed drug has received full FDA approval. AI-native biotechs actually do better than the industry average early on, clearing Phase 1 at roughly 80 to 90 percent versus an industry average near 40 to 65 percent. Then Phase 2 hits, and the success rate drops to about 40 percent. That's still ahead of the roughly 29 percent industry average, but it's the same wall every drug developer runs into: does it work safely in a large, diverse group of actual patients.
A faster model that predicts protein structure doesn't answer that question. Only a body does. That's the honest limit sitting underneath Anthropic's wet lab announcement. Computational speed helps researchers generate more candidates and rule out bad ones faster. It does not shortcut the part of drug development that has always been slow: watching what a compound does inside a living system over months or years.
Nexairi Dispatch
Get the next AI move before it turns into a workflow problem.
Join the free newsletter for concise AI news, practical checklists, and the decisions practitioners need to make next.
Free. No spam. You will also get the Nexairi Dispatch.
Why is Anthropic worried about this now?
Because the same capability jump that makes Claude useful in a lab also makes it more useful to someone trying to cause harm.
Anthropic's own threat report, released this September, said something no major AI lab had said publicly before. Its older models were confidently below the skill level needed to meaningfully help someone develop a biological weapon. Its newest models, the report says, can no longer be assumed to sit below that line. Anthropic disclosed it had already identified and blocked five separate attempts to misuse Claude for research that could support bioweapons work, including a gain-of-function study connected to a military-linked research institute.
That admission is why the Life Sciences Verification Program exists. Its default, Fable-class models still block a wide range of professional biology and drug-development queries, on purpose. Vetted researchers who pass verification can apply for deeper access, either "Standard Use," which unlocks basic science and clinical work while keeping other safety checks in place, or "High-risk Use," a narrower grant renewed every six months for specific high-stakes projects. Anthropic also shifted its monitoring approach for this program from blocking requests in real time to reviewing flagged activity afterward, holding that data for 30 days.
So the lab and the lockdown are not two separate stories. They're the same story told from opposite ends. Anthropic wants Claude doing more real biology work. It also just told the world it can no longer promise that capability stays safely out of the wrong hands by default.
What this means if you're not a biologist
You don't need to run a lab to feel this. If you work in pharma, biotech, health IT, or manage vendor risk for a company touching life sciences, watch the access model. Anthropic drawing a hard line between what a general Claude user can ask and what a verified researcher can ask is a pattern other AI labs in sensitive domains, like chemistry or cybersecurity, will likely copy. Expect a verification step to become normal.
Second, don't mistake a faster model for a faster industry. The venture money chasing AI biopharma is betting that computational speed eventually compounds into approved drugs. The Phase 2 numbers say that bet hasn't paid off yet, and won't just because Anthropic's models got quicker at folding proteins. Biology still has to answer to biology. The question worth asking any AI-biotech vendor pitching you right now isn't how fast their model is. It's how many of their predictions have survived contact with an actual patient.
The useful move here isn't picking a side on whether Anthropic's wet lab is exciting or alarming. It's watching whether the Life Sciences Verification Program's offline-monitoring approach actually catches misuse after the fact as reliably as the old real-time blocks caught it before. That's the detail that will tell you whether this access model was the right trade, or just a faster front door with a slower alarm.
Sources
- Anthropic: How Claude Is Uplifting Biomolecular Modeling
- Anthropic: Introducing the Life Sciences Verification Program
- Anthropic: Expanding Our Support for Scientists
- CNN: Anthropic Says It Blocked Misuse of AI That Could Have Supported Biological Weapons
- Axios: AI Is Not Yet Driving Drug Development
- Techstrong.ai: Anthropic Quietly Opens Bay Area Wet Lab
Related Articles on Nexairi
Jim Smart is the founder and editor in chief of Nexairi. A Business Intelligence Developer with experience building data systems for Verizon, U.S. Army operations, and enterprise finance teams, Jim spent years turning complex data into decisions that executives could act on — dashboards, forecasting models, and automation pipelines across telecom and government contracting. He founded Nexairi to apply that same clarity to AI: making emerging technology understandable and actionable for the operators, accountants, and business owners who need it most. Jim holds GenAI certifications from the University of South Florida Bellini College of AI and completed Springboard's Data Science Career Track.
