Anthropic CEO Dario Amodei: Claude Can Autonomously Operate Experimental Equipment
Anthropic CEO Dario Amodei stated that Claude can be used to autonomously control laboratory equipment and conduct biological experiments, advancing models from analytical hypotheses to physical world operations.
The company has established a wet lab in the San Francisco Bay Area, with life sciences head Eric Kauderer-Abrams confirming to Reuters that biological validation is still conducted in real experiments, with the company synchronously working with both its own facilities and external partners. The goal is for Claude to command robotic units to complete experiments with limited human intervention. The company stated it will not take the project to clinical trials to avoid direct competition with pharmaceutical clients.
The supporting interface is the Model Hardware Standard launched in August, allowing agents to drive microscopes, liquid handlers, lasers, and robotic arms with code. Test subjects include Howard Hughes Medical Institute, Carnegie Mellon University, Genentech, and quantum computing company QuEra. Jonah Cool mentioned that complex microscopes with hundreds of mirrors, motors, and lasers, which previously required expert operation, can now be directly controlled by Claude through dialogue.
There are already experimental records: Alek Kemeny observed Claude controlling mirrors and lasers in a neuroscience experiment, identifying the alveus structure in live brain tissue; Carnegie Mellon used liquid handlers, plate readers, robotic arms, and surveillance cameras to create dose-response curves, claiming speeds nearly three times faster; in Genentech's BCA protein assay, Claude mistakenly identified bubbles as software faults and added foam, requiring human intervention. The company stated that from bare metal to completing dilution curves, including one autonomous rerun, takes about eight hours, while vendor turnkey solutions usually take weeks.
On the software side, there is the Claude Science workstation, connecting over 60 genomics, proteomics, and cheminformatics databases, capable of running tasks on the lab's own clusters or HPC. In external evaluations, Adaptyv Bio and Twist Bioscience validated Claude's designed binding proteins for 15 targets, with 14 successful; after adding two prompts to the original documents from contract labs, Claude provided NMR and mass spectrometry results in 23 and 19 minutes, with purity readings of 96.4% compared to the lab's 96.33%.
In market mechanisms, the workflow entry points are purchased by pharmaceutical companies, academic institutions, and instrument vendors, while the model invocation and equipment orchestration are sold. Funding has shifted from traditional CRO scheduling and instrument integrators' custom software to an agent layer billed per experimental cycle. Beneficiaries include Anthropic, Nvidia's BioNeMo toolchain, and Gigs-style embedded hardware vendors; those under pressure are laboratory automation vendors relying on manual operation and proprietary software to bind clients, as well as small biotechs outsourcing discovery phases to slow-cycle CROs. The event-driven clarity: the establishment of wet labs transforms "models that can write protocols" into "models that can operate at the push of a button."
Amodei has previously written that his father passed away years before a therapy became available; he also mentioned that artificial intelligence could compress decades of biological progress within 5 to 10 years. The company simultaneously warns of the risks of bioweapons misuse and claims to have intercepted abuses of Claude related to bioweapons. Job postings specify the goal of accelerating progress in life sciences by an order of magnitude.
Source: Public Information
ABAB AI Insight
Dario Amodei, a PhD with a background in biophysics, transitioned from OpenAI to establish Anthropic as a safety narrative company, yet has consistently developed products into industry operating systems: Claude Code for software engineering, Claude Science for research workstations, and acquiring the stealth biotech Coefficient Bio to bring wet experiments in-house. This mirrors Demis Hassabis's approach of using AlphaFold for structural prediction and Isomorphic for drug development, but Anthropic chose to first build the orchestration layer rather than specialized structural models. He has publicly noted that his father died from a disease that had no drug for years, giving a personal motivation to the life sciences line, along with a self-imposed limitation of "only reaching preclinical" for pharmaceutical clients.
The capital path is: computational power subscriptions sustain models, which then drive CyBio Felix liquid handlers, Varioskan LUX plate readers, and robotic arms, feeding curve readings back to the model to decide whether to rerun. Funding has shifted from laboratory headcounts and integration projects to tokens and orchestration interfaces. The strategic motive is to occupy the feedback loop in the discovery phase—whoever first connects "design—liquid transfer—reading—parameter adjustment" will increase the migration costs for pharmaceutical companies. Novo Nordisk has entered into collaborative discovery, indicating that resource transfer is not just for in-house laboratories but also integrates big pharma's targets and data into the same agent framework.
Analogous examples include Tesla connecting visual models to production line robots, and Recursion and Insitro using automated wet experiments to feed their own models. The industry phase has shifted from "model-assisted protocol writing" to "model operational oversight." Google DeepMind still focuses on foundational models and structural predictions, OpenAI leans towards general assistants, while Anthropic is competing for the laboratory operating system layer, positioning itself closer to Thermo Fisher's software layer rather than becoming another virtual pharmaceutical company.
The structural judgment is that technological substitution overlaps with industry chain reconstruction. The replacement is not of scientists' judgments, but of operational oversight, interfacing with proprietary instrument software, and transferring 96-well plates from processing instruments to plate readers. The mechanism at play is the long-term fragmentation of instrument interfaces, which can take weeks to integrate; once a unified driving layer emerges, the marginal experimental cost shifts from "scheduling + engineers" to "electricity + tokens." Pricing power shifts from companies selling instruments and labor to those selling orchestration models that can read errors and decide to rerun. Wet laboratories are not intended to become pharmaceutical companies themselves, but to ensure that this orchestration passes through real bubbles and real focus errors.
ABAB News · Cognitive Laws
The feedback loop is not in the papers, but in the buttons that can be pressed.
Whoever first understands instrument error messages will control the rhythm of experiments.
Safety narratives are sold to regulators, operating systems are sold to clients.