A Model Company Built a Biology Lab. Here Is What It Is Testing.

Reuters reported that Anthropic built a biology lab in the San Francisco Bay Area. Eric Kauderer-Abrams, the company's head of life sciences, confirmed it. His reason: biological research ultimately has to be tested through physical experiments. Anthropic started doing some of that work internally and kept working with outside labs.

A few days later, Anthropic published an early result. About 950 Claude agents analyzed more than 200,000 reverse transcriptase-related sequences. They found an enzyme system with features that resemble CRISPR repeat sequences. The team then confirmed part of the picture in the lab: the repeats can express short RNAs. Their function and possible use remain unclear. The system is called ART, for array-associated reverse transcriptases.

Anthropic has touched life sciences before. Over the past year it launched products for scientists, worked with biology institutes and drug companies, and had Claude help with protein design, call specialized models, and explore control of lab instruments. The pattern points to a goal beyond a single research tool.

The obvious question is why a general model company would run a biology lab.

Life science does not lack tools. It lacks someone to connect them.

Protein design already has models for structure generation, sequence design, and binding prediction. Researchers use them, but they still choose the model, set the parameters, and decide which computational hits go to the bench. Each tool handles one step. A person handles the space between steps.

Anthropic first aimed at that space. In October 2025 it launched Claude for Life Sciences, which let Claude connect to Benchling, 10x Genomics, and other platforms, read lab records, process sequencing data, and help draft research plans. Claude Science came later and put scientific databases, specialized models, and computing resources into one workbench.

Anthropic did not build a full biology software suite. It lets Claude call existing tools and organize the workflow. Claude does not perform every calculation. It picks tools for the task and helps scientists keep the process connected. That position lets Anthropic reach more stages, including clinical trial management and regulatory filings.

Connecting tools is one thing. Doing the work reliably is another. Researchers care about whether Claude can finish a task and produce something they can check.

Anthropic has worked with the Allen Institute and the Howard Hughes Medical Institute on complex biological data, experimental design, and hypothesis generation. It also launched BioMysteryBench, a set of 99 tasks drawn from biological data. The test asks whether a model can call tools, analyze data, and return verifiable results on its own.

A protein design experiment made the point more directly. Claude did not replace RFdiffusion or other specialized models. It organized them for design, parameter tuning, and candidate screening. Anthropic sent 1,320 designs to outside labs. Labs confirmed that 354 bound target proteins, a hit rate of about 26.8%. Of the 15 targets in the final count, 14 received at least one successful design.

Those results are far from a drug. They do show that Claude can use existing models to finish a complex research task and produce candidates that survive lab testing. Anthropic does not need its own model in every biology subfield. It wants Claude to be the common entry point for the tools.

Protein design has a clear target. Scientists know the kind of molecule they want, and Claude looks for designs that fit. Basic biology is looser. Scientists sometimes do not know what they are looking for.

From completing tasks to finding questions to study

Nature holds many proteins and gene sequences with unknown functions. Sequencing data keeps growing. Finding a system with an unusual function still takes long rounds of comparison. This spring Anthropic formed a life sciences team with backgrounds in computational biology, CRISPR systems, and novel enzymes. The team runs basic research. It uses Claude to search DNA data for unknown proteins and systems, then tests the findings in the lab.

ART was one early project. Researchers gave Claude a direction: find unknown reverse transcriptase systems. About 950 agents then analyzed more than 200,000 sequences in parallel. Claude found regularly spaced DNA repeats near a class of reverse transcriptase genes. The structure looked similar to CRISPR systems. Follow-up lab work showed that the repeats can express short RNAs. The biological function remains unknown.

The study does not prove that Claude can make a scientific discovery alone. People set the direction, reviewed candidates, and ran the experiments. Claude's role changed, though. In protein design it selected tools and screened candidates against a known goal. In ART it helped find a question that might be worth pursuing.

That change creates a new problem. Claude can analyze large datasets fast, and it can produce more candidates than a lab can test. In ART, one research run produced hundreds or even thousands of candidate reports. No scientist can run all of them. The hard question is which leads to test first. Anthropic has started teaching Claude from how researchers choose questions, but that work is early.

The lab bench sets a harder limit. A model can generate molecules or hypotheses in minutes. A protein still has to show the expected function in a physical sample. A newly found enzyme still has to show what it does in a cell. Programming can run code, read the error, and try again. Biology needs samples, instruments, trained people, and time. If validation lags, a larger pile of candidates does not help.

The case for its own lab

Anthropic could already get lab work through partners. Outside labs validated the protein design study. The company also runs protein design competitions with Adaptyv Bio, Twist Bioscience, and others, with plans to test more than 5,000 screened designs and publish the results. Partners bring established lab capacity. Anthropic does not need to build every facility.

Claude's move into open-ended research changes the demand. Testing a batch of finished designs is not the same as probing an unknown biological system. The second kind of work often needs a loop: run an experiment, reanalyze the data, adjust the hypothesis, then choose the next experiment. The workflow is not fixed. Computational and lab staff have to work close together.

Kauderer-Abrams told Reuters that an internal lab would speed some work and give the team direct lab experience. Anthropic will keep working with outside labs and choose the right setup for each task.

The lab does not mean Anthropic will run every life science experiment itself. It lets the company join some experiments and see where Claude helps and where it fails in a working lab. The lab fills a validation loop. A model proposes. An experiment tests. The result goes back to the model and the researchers.

Anthropic is also working on two support problems: making lab equipment easier for AI to use and cutting the computing cost of biology models.

For equipment, Anthropic worked with HHMI's Janelia Research Campus on the Model Hardware Standard. The standard gives AI agents one interface for microscopes, liquid handlers, robotic arms, and other instruments. That reduces software integration work between machines. Claude Science connects databases, models, and tools inside a computer. MHS extends that link to lab hardware.

For computing, Anthropic used Claude to optimize more than 30 open-source biomolecular models. In company tests, the models ran about four times faster on average. The work targets existing model implementations and can lower computing needs for some research tasks. MHS is still a research preview, and the speed gains depend on the task and the environment. Both efforts point the same way: Claude should understand a research task and get easier access to the resources that task needs.

Hiring follows the same path. Anthropic acquired the Coefficient Bio team, which has a computational drug discovery background. John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, said he is joining. Anthropic is also hiring in computational biology, experimental biology, and chemistry. For a general model company, these people improve Claude's research work. They also let Anthropic run some biology research itself.

Anthropic does not plan to become a traditional drug company. Kauderer-Abrams told Reuters that the company is focused on rare diseases, neglected diseases, and problems that have been hard to drug. It does not plan to run commercial clinical trials or compete directly with biotech companies that sell drugs.

The current plan is to support research institutions and drug companies through Claude. Anthropic will not carry a drug from idea to market. The internal lab helps the team test model-generated hypotheses and gain lab experience.

Back to the opening question

The opening question was why a model company would build a biology lab.

Line up the past year: Claude connected to research data and specialized models, took part in protein design, probed unknown biological systems, and now has internal lab capacity. Each step pushes Claude deeper into life science research. Anthropic does not need its own foundation model in every biology subfield. It wants to control the entry point that links models, data, and lab workflows.

That entry point faces concrete tests. Can lab validation keep up with the candidates AI produces? Can the loop run when biology experiments are slow and costly? Who owns safety and responsibility when AI connects to lab equipment? Will research institutions and drug companies keep paying for this kind of infrastructure?

One detail from the Reuters report matters. The lab exists. The ART repeats can express short RNAs. Their function is unknown. That gap between a working bench and a working answer is where Anthropic sits now.