Anthropic Confirms Wet Lab as Claude Moves Into Physical Drug Research

Anthropic has moved beyond computer simulation: the company quietly established a physical wet lab in the San Francisco Bay Area, and its head of life sciences, Eric Kauderer-Abrams, confirmed the facility to Reuters on September 18, 2026. The disclosure is the clearest signal yet that the maker of Claude is no longer positioning itself as an AI platform for pharmaceutical customers alone — it is now operating scientific infrastructure of the kind more typically found at biotech startups. Life sciences has become one of Anthropic's largest investment areas by headcount and resources, Kauderer-Abrams said, and the company is testing whether Claude can guide robotic systems through physical experiments with minimal human intervention.

"We believe that to do biology, the final test is still and will be for a while in real lab work," Kauderer-Abrams told Reuters. "We absolutely are doing that today." An Anthropic spokesperson added a notable qualifier: the existing lab is not specifically for drug discovery. The company declined to say what, precisely, the facility is being used for. Kauderer-Abrams also told Reuters that Anthropic's approach mirrors standard biotech practice — combining its own laboratory work with external partner collaborations — and was direct about the company's limits: "We're not competing with pharma and biotech companies that make their business in bringing drugs to market."

Why the In Silico-to-Wet Lab Transition Matters

The distinction between in silico and wet lab biology is substantive. Computational methods predict protein folding, simulate molecular interactions, and screen candidate compounds. A wet lab tests those predictions against physical reality — real enzymes, real cells, real chemical reactions. Every AI drug discovery company eventually confronts the gap between a model's output and what happens in a test tube. Isomorphic Labs, Alphabet's dedicated drug-discovery subsidiary, has spent years bridging that gap and still expects its first AI-designed molecules to reach clinical trials only by the end of 2026 — a deadline it already pushed back by one year, after CEO Demis Hassabis announced the delay at the World Economic Forum in January.

Anthropic is not claiming to be at that stage. Kauderer-Abrams described the lab automation effort as in "very early innings," and the company said explicitly that it is not running clinical trials. The wet lab gives Anthropic what it previously lacked: a physical feedback loop. When Claude makes a prediction about a molecular interaction or an experimental protocol, Anthropic can now test whether that prediction survives contact with actual biology — and use those results to improve both the model and the workflow.

Claude Agents in the Lab: The Model Hardware Standard

The infrastructure connecting Claude to physical instruments was already partly visible before the wet lab story broke. On August 27, 2026, Anthropic previewed the Model Hardware Standard (MHS), a specification that lets AI agents discover and operate laboratory equipment — microscopes, liquid handlers, robotic arms, plate readers — through a unified driver layer. Each MHS driver translates between the operating system and the hardware device using simple read-and-write commands, and stores device characteristics such as weight, safety limits, and adjustable parameters that previously existed only in paper manuals. Safety limits are enforced at the driver layer, below the AI agent, so an agent cannot instruct a device to exceed its physical boundaries regardless of what the model might otherwise attempt.

Read more: Anthropic's Model Hardware Standard for AI lab agents

Genentech tested MHS on a protein assay workflow spanning a liquid handler, robotic arm, and plate reader; Claude converged on optimal transfer parameters without human intervention at each step. QuEra Computing used the same standard to automate quantum laser-lock recovery: a bespoke script built over months by a four-person engineering team succeeded about 58 percent of the time. After an overnight MHS-based agent run, the success rate jumped to 99.3 percent across 700 trials. Anthropic acknowledged in that MHS announcement that Claude "lacks deep physical intuition regarding chemical and biological phenomena" — at Genentech, researchers had to guide Claude to recognize that liquid foaming was a physical failure rather than a software bug. MHS currently works only with equipment that exposes a programmable control interface.

One Stack, End to End

The wet lab sits at the end of a biology infrastructure build that began publicly in October 2025 and accelerated sharply through mid-2026. In April 2026, Anthropic acquired Coefficient Bio — a stealth startup of fewer than ten people, most of them former Genentech computational biology researchers — for approximately $400 million in stock. The founders, Samuel Stanton and Nathan C. Frey, came from Prescient Design, Genentech's computational drug discovery unit, bringing direct experience in AI-based molecular design. On June 30, 2026, Anthropic launched Claude Science, a research workbench pre-configured with more than 60 scientific databases covering genomics, proteomics, structural biology, and cheminformatics, running on existing Claude models. Novartis CEO Vas Narasimhan joined the Anthropic board the same day. Partnerships with Genentech, Bristol Myers Squibb, and Novo Nordisk followed, the last formalizing a collaboration on September 16, two days before the wet lab story appeared.

Taken together — Coefficient Bio's molecular expertise, Claude Science's research workflow, MHS's hardware interface, and the wet lab's physical execution layer — these form a vertically integrated biology stack. No other frontier AI lab has assembled all four layers simultaneously. OpenAI's GPT-Rosalind, a domain-specific model for biological reasoning launched in April 2026, and Google's Gemini for Science, which integrates AlphaFold and AlphaGenome with its own database set, compete at the model and workflow levels. Neither has published a hardware standard comparable to MHS, and neither has confirmed physical laboratory operations of its own.

Targeting Undruggable Biology

Kauderer-Abrams identified a specific category of diseases and molecules as Anthropic's focus: conditions long considered "undruggable" — targets without the well-defined binding pockets that allow conventional small-molecule drugs to attach. Many transcription factors, intrinsically disordered proteins, and protein-protein interaction interfaces fall here, including oncogenic targets that pharmaceutical companies have wanted to reach for decades but lacked the chemistry to do so.

Bispecific and trispecific antibodies are one technical route toward these targets. Unlike standard antibodies that bind a single molecular site, bispecific antibodies are engineered to simultaneously engage two distinct targets or two sites on the same target — for example, recruiting a T-cell immune effector into proximity with a tumor cell that carries a cancer-associated antigen. The combinatorial design space is large, and each candidate requires computational screening followed by wet lab confirmation. That combination is precisely the loop Anthropic is now equipped to close.

Anthropic has not disclosed what specific diseases it is working on or what experimental results the wet lab has produced. Kauderer-Abrams said the company's focus is on needs the commercial market has overlooked. Finding a promising candidate molecule is only the beginning; animal studies, clinical trials, and regulatory review follow before any treatment could reach patients — a process that typically spans years and that Anthropic has explicitly not committed to undertaking.

Biosecurity Questions and Data Conflicts

The wet lab disclosure arrives alongside escalating biosecurity debate. According to Anthropic's published safety reports, the company disclosed several biology-related cases between late 2025 and mid-2026 in which users attempted to use Claude to access sensitive biological research through bypassed safety controls. No case produced a synthesized pathogen or real-world incident. Experts remain divided: some biosecurity researchers argue that physical bottlenecks — access to equipment, precursor materials, and specialist knowledge — make near-term AI-enabled bioweapon development unlikely; others argue that frontier models provide meaningful capability to non-expert actors. The Gizmodo report on the wet lab disclosure noted Anthropic had released a safety report on attempted biological misuse just one week before the wet lab story emerged, and Kauderer-Abrams addressed the tension directly: "we're always balancing these things," he told Reuters, adding that deployment is managed carefully to reduce risk. US Congress has held hearings on AI biosecurity, and federal oversight of dual-use biological research has tightened in 2026.

A separate tension concerns competitive data. Anthropic now provides AI services simultaneously to Genentech, Bristol Myers Squibb, and Novo Nordisk — pharmaceutical companies that compete directly in multiple therapeutic areas. The company says it isolates customer data between partners, but a vendor that runs its own drug program while serving multiple competing pharmaceutical clients occupies a structurally unusual position in the industry.

The Clock and What Comes Next

Isomorphic Labs remains the most advanced publicly disclosed AI drug program. Founded in 2021, backed by Alphabet, and building on AlphaFold protein structure prediction, Isomorphic targets its first clinical trials by the end of 2026 — already one year later than its original promise. Anthropic's wet lab was confirmed this week. The Coefficient Bio team joined in April. Claude Science launched three months ago. On a realistic drug development timeline, Anthropic is years from any clinical-stage program, and Kauderer-Abrams has said nothing to suggest otherwise.

What Anthropic has built is a feedback architecture that no other frontier AI lab currently operates in full: a chain from model reasoning through software workflow to hardware interface to physical experiment. Whether that chain produces scientifically meaningful results is something only undisclosed wet lab data can answer. The company's approaching IPO — investor accounts expect a possible $2 trillion valuation, with a market debut potentially as early as October — will put increasing pressure on the life sciences division to show outcomes rather than infrastructure.

Read more: Anthropic targets record IPO as Claude revenue surpasses $65 billion

For pharmaceutical companies evaluating Claude Science against OpenAI and Google's competing platforms, the wet lab is either the differentiator that tips the decision or a distraction from the question of which model is most useful at the bench. That distinction will become clearer as Anthropic moves from announcing the loop to reporting what it produces.