Anthropic confirmed the opening of its own wet lab, a facility for physical biological experiments, in the San Francisco Bay Area, according to a Reuters report on September 18, 2026. The company is already conducting real laboratory work and, according to Eric Cauderer-Abrams, head of the life sciences division, is in the "earliest stages" of automating experiments: Claude is intended to control laboratory robots with minimal human intervention and mandatory human oversight. The company declined to disclose the size of the lab, its opening date, and its biosafety level.

image

What happened

The creation of a wet lab in the San Francisco Bay Area was reported on September 18, 2026, by Reuters and CNBC, citing confirmation from Anthropic itself. Eric Cauderer-Abrams, who leads the life sciences division, stated that the company is already conducting real laboratory work, not just preparing for it. The automation of experiments was described as being in its "earliest stages": Claude is intended to take over the control of laboratory robots while remaining under mandatory human oversight. An Anthropic representative clarified that the lab was created "not specifically for drug discovery" but declined to name its size, opening date, and biosafety level. There are no published results from the lab's work, benchmarks, or descriptions of automation protocols in open sources.

Context

The lab is part of Anthropic's broader expansion into life sciences: the company acquired the biotech startup Coefficient Bio for approximately $400 million in stock, added Novartis CEO Vas Narasimhan to its board of directors, launched the Claude Science product, and is working with pharmaceutical clients Genentech, Bristol Myers Squibb, and Novo Nordisk. Until this point, the work of large language model creators with biotech was limited to "in silico" experiments, i.e., computational modeling without a physical laboratory. Before Anthropic, wet labs were built by pharmaceutical companies and Isomorphic Labs, an Alphabet project that has already postponed its clinical target to the end of 2026. The difference in this new step is that the physical experimental infrastructure belongs to the creator of the model itself, not to a pharmaceutical company or research division.

Why this matters for the industry

For the AI industry, the main shift is that a language model is transforming from a data analyst into an executor of real laboratory protocols, and the combination of a wet lab, Claude, and laboratory robots sets a template for a new class of scenarios, an "agent-orchestrator of a physical laboratory" with a human in the loop. Vertical integration makes the model creator the owner of the source of experimental data, which changes the familiar cycle of model evaluation and sets a competitive benchmark for OpenAI, Google, and biotech startups. For startups, the signal is twofold: a market for lab automation around Claude is opening up, but the risk is also growing that the vendor itself is moving down the stack and competing with its own partners. Right now, there is nothing to deploy: the lab is internal Anthropic infrastructure without an API, pricing, metrics, or public evals, and the product available to third-party users, Claude Science, is currently designed for analytical work. If the automation is confirmed, model creators will close the full cycle of "hypothesis, experiment, data," obtain closed, vendor-owned experimental datasets, and shift model evaluation from text benchmarks to physical ones, and in the long-term scenario, a market for middleware between LLM agents and laboratory equipment may form.

Why this matters for users

For the reader, this is the first case where the creator of one of the leading LLMs publicly owns its own biology lab and claims to control robots through Claude, so it is now possible to track how quickly "AI automation of science" will reach reproducible real experiments through the company's public steps. The news has no immediate practical benefit: the lab is closed to external access, and of Anthropic's products in this area, only Claude Science is currently available for analytical work, not for controlling equipment. A separate reason for attention is biosafety: Anthropic itself has reported cases where its models could be used to develop biological weapons, and now the company is maintaining a physical biology lab without disclosing its level of protection.

What is still unknown / limitations

The claim that Claude is already controlling laboratory robots is stronger than the available data: sources state that the automation is in its "earliest stages," so the fact of the lab's existence and the work being conducted in it are confirmed, but not a working autonomous control loop. The stated ability of the model to control laboratory equipment is not accompanied by a description of the agent architecture, evaluation protocols, or a human-in-the-loop interface. The phrasing "not specifically for drug discovery" leaves the main question about the object's purpose open: perhaps the lab is used to generate its own training and evaluation data for science tasks. The demonstrations and pharma partner cases expected in the coming months are an interpretation of prospects, not confirmed facts.

Sources

Author

Look at AI, editorial team