Anthropic Takes Drug Discovery Off the Screen and Into the Lab
Anthropic is building a biology laboratory of its own, a move that shifts the company's life-sciences work from purely computational modeling into physical experimentation. Reuters reports that the facility will sit in the San Francisco area, and that the central premise is unusual: Claude, Anthropic's family of AI models, will orchestrate robotic lab equipment through drug experiments with as little human intervention as possible.
The distinction matters. Most AI-driven drug discovery today happens inside software — generating candidate molecules, predicting binding affinity, ranking compounds against a target. Those predictions still have to be validated at a bench, and that validation is where timelines stretch and costs accumulate. By standing up its own lab, Anthropic is signaling that it wants a tighter loop between the model that proposes an experiment and the hardware that runs it.
Why Physical Experiments Are the Hard Part
Bench science is slow, sequential and unforgiving. Protocols run in steps, each dependent on the last, and a single contaminated plate can invalidate a week of work. The pitch for automation is throughput and reproducibility: robots do not get tired or lose focus, and an AI system can log every parameter of every run.
But biology is noisy. Assays fail, reagents degrade, and instruments drift out of calibration. A model trained largely on literature and simulation has limited intuition for what happens when a reagent misbehaves at 4 p.m. on a Friday. That is likely why Anthropic frames the ambition as minimal human involvement rather than none at all, and why the company says people remain part of the process for safety reasons.
Two Tools Already Shipped to Enable the Work
Anthropic has not jumped straight to hardware without preparation. It has already released two pieces of infrastructure designed to make model-driven experimentation possible.
- Claude Science — an AI workspace built specifically for researchers, giving scientists a structured environment in which to work through scientific problems with the model.
- Model Hardware Standard — a specification intended to let AI models control laboratory equipment, addressing the translation layer between a model's instructions and the instruments sitting on a bench.
The hardware standard may be the more telling of the two. Lab equipment has historically been a patchwork of vendor-specific software with limited interoperability. Handing reliable control to a model requires agreed interfaces, error handling and state reporting, otherwise an instruction like "run this assay" has nothing concrete to bind to.
Human Oversight Stays in the Loop
Anthropic is explicit that human oversight remains a requirement. That is not just regulatory caution. An AI system that can both design and execute experiments can also make mistakes at scale — a mis-specified protocol run hundreds of times produces hundreds of wasted runs, and a subtle error in a robotic liquid-handling step can quietly corrupt a dataset rather than fail loudly.
Keeping people in the loop also matters for interpretation. Deciding that an unexpected result is a discovery rather than an artifact is still a judgment call that depends on context a model may not possess. The design question for Anthropic is where that judgment sits in the workflow, and how much of the experimental cycle can run unattended before a human signs off.
The Prize: Diseases Long Called 'Undruggable'
Eric Kauderer-Abrams, who runs Anthropic's life sciences division, points to conditions that have resisted treatment entirely because no viable therapy exists at all. His bet is that AI can accelerate the design of complex antibodies capable of hitting multiple targets at once — a class of molecule that is notoriously difficult to engineer by conventional means.
Multi-target antibodies are attractive because many diseases involve more than one biological pathway, and hitting several at once can be more effective than blocking a single node. The engineering problem is that a molecule binding two targets must stay stable, manufacturable and specific; the design space is vast and mostly empty. Kauderer-Abrams sees AI as a way to search that space faster than iterative human design allows.
Built on a Year of Biotech Moves
The lab does not arrive in isolation. Earlier this year, in April, Anthropic acquired the startup Coefficient Bio for roughly $400 million. Around the same period, Novartis CEO Vas Narasimhan joined Anthropic's board, adding pharmaceutical-industry perspective at the governance level.
Read together, those moves sketch a deliberate strategy: acquire biological expertise, bring in pharma leadership, ship tooling that researchers can actually use, and then build physical capacity to generate proprietary data. Each step compounds the last. A lab that feeds its own experimental results back into model development would give Anthropic something few AI companies possess — a closed loop between prediction and physical ground truth.
No Clinical Trials, For Now
One boundary Anthropic appears to be drawing: it will not run its own clinical trials at this stage, according to the report. The rationale is pragmatic and diplomatic. Human trials are the most expensive, slowest and most heavily regulated phase of drug development, and moving into them would put the company in direct competition with the pharmaceutical partners it plausibly wants to work alongside.
Staying upstream — in discovery, design and preclinical validation — lets Anthropic sell capability rather than compete for the same pipeline. It also keeps the company out of the liability and compliance footprint that comes with running studies in patients.
What to Watch
Several questions will determine whether this becomes a genuine shift in how drugs are discovered or an expensive demonstration. How much of the experimental cycle is actually autonomous, and how often does a human intervene? Will results be published, or is the data destined to stay proprietary? Does the Model Hardware Standard get adopted by instrument makers beyond Anthropic's own benches?
For now, the significance is directional. An AI lab that can design molecules and then physically test them closes a gap that has limited the field for years. Whether Anthropic's robots can make that loop faster and more reliable than a well-run human lab remains the experiment that matters most.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
