Anthropic’s Vetted Biology AI Program and Bay Area Wet Lab, Explained

Anthropic’s September 2026 life-sciences push has two connected parts: a verification program that gives approved organizations more permissive access to Claude for biology-related work, and a reported San Francisco Bay Area wet lab for testing work in the physical world. Together, they signal an effort to make Claude useful across computational research and experimental science while retaining controls around dual-use biological capabilities. 617

What the Life Sciences Verification Program does

The beta Life Sciences Verification Program (LSVP) opened applications to life-science teams and institutions after Anthropic had already onboarded dozens of organizations through early access. Participants are vetted for research credentials, security practices, and ethical-research oversight. 6

Approved users can access Claude Mythos, Opus, and Sonnet with biology safeguards designed to be more permissive for legitimate professional work than ordinary consumer access. Anthropic positions the program for research biology, drug discovery, clinical development, manufacturing, and related life-science workflows. 6

The beta initially covers teams and institutions rather than individual subscribers. Anthropic said it expects to expand the program to individual Pro and Max plans over time, but did not provide a timetable. 6

Standard Use vs. High-risk Use

LSVP separates access into two grant types:

• Standard Use is intended for routine, broad team use in normal life-science work. It covers Mythos 5.1, Opus 5, and Sonnet 5, renews annually, and is designed to carry forward to future models. 6
• High-risk Use is a project-specific add-on for work that Standard Use would block. It removes life-science blocking safeguards for the approved project, renews every six months, and initially applies to Opus 5 and Sonnet 5. Anthropic said broader high-risk access to Mythos remained limited while it coordinated with the U.S. government. 6

This is not unrestricted model availability. The program is structured around verified organizations, specified use cases, and grants that can be renewed or reviewed.

How the safeguards work

Anthropic identified three main risk categories for the program: compromised accounts or malware-driven takeover, insiders diverting legitimate access, and autonomous agents—including long-running or swarm-like systems—taking unsafe actions. 6

Rather than relying mainly on real-time refusals, the program links access to declared use cases and uses offline monitoring to look for suspicious patterns across requests and sessions. Where activity is flagged, cases can be escalated to an organization’s administrators under agreed response timelines. 6

Traffic associated with this monitoring is retained for 30 days. Anthropic says that data is compartmentalized, excluded from model training, and unavailable to its life-sciences research teams. 6

Anthropic also said it had observed increasingly sophisticated attempts to misuse AI in ways that could support biological-weapons development. The supplied reporting does not identify pathogens, actors, or methods, so those details should not be inferred. 6

Why Anthropic built a wet lab

Reuters reported that Anthropic had set up a wet lab in the San Francisco Bay Area for physical biology work, and Anthropic life-sciences head Eric Kauderer-Abrams confirmed the lab’s existence. The company said it would perform some work internally and some with outside partners; a spokesperson clarified that the facility was not specifically a drug-discovery lab. 17

The strategic value is feedback from experiments rather than simulations alone. Anthropic is reportedly exploring whether Claude can direct robotic systems that perform laboratory experiments with limited human intervention, while maintaining human oversight. 17

That does not mean Claude is independently running a drug-development pipeline. It means Anthropic is building capacity to test whether AI-generated scientific ideas hold up under physical experimental conditions.

The model-optimization claim

Separately, Anthropic reported that Claude optimized more than 30 open-source biomolecular models in just under four weeks. According to the company, the optimized models ran roughly 4× faster on average and used less memory; it also described a low-memory mode for accurate prediction of biomolecular systems larger than 10,000 tokens on a single NVIDIA GPU node. 42

These results point to a practical use of AI in science: improving the computational tools researchers already rely on for tasks such as biomolecular prediction and design. But they remain Anthropic’s own reported findings, and the supplied evidence does not substantiate a specific protein-design cost comparison or an Apache 2.0 release claim.

The larger commercial and scientific strategy

The program and laboratory sit alongside Anthropic’s broader life-sciences initiative, including Claude Science and an internal drug-discovery effort focused on diseases the company characterizes as neglected or commercially unattractive to traditional pharmaceutical development. 27

Reuters reported that this expansion came as Anthropic was preparing for a potential IPO and working with major pharmaceutical companies, including Roche’s Genentech, Bristol Myers Squibb, and Novo Nordisk. 17

Anthropic’s stated line is to support the industry rather than directly compete across the full pharmaceutical value chain: it has framed its own preclinical work around neglected needs, not bringing mainstream drugs to market or running clinical trials. 1727

Bottom line

LSVP is Anthropic’s attempt to solve a difficult access problem: give qualified scientists fewer unnecessary barriers in legitimate biology work without treating high-capability bio AI as a general-purpose public tool. The Bay Area wet lab extends that strategy from software into experimental validation. Whether this produces faster scientific progress will depend not only on model capability, but also on the quality of the experiments, partner oversight, and the safeguards governing how the systems are used. 617