On September 17, 2026, Anthropic announced the launch of the Life Sciences Verification Program (LSVP), a pilot program that enables life sciences professionals to use the Mythos, Opus, and Sonnet models under more flexible controls for biology-related tasks. The company says it has already enrolled dozens of institutions in early access and has begun accepting applications from a broader segment of the life sciences community.
The program targets teams and institutions including academic laboratories, startups, and pharmaceutical companies, and covers use cases such as drug discovery, biological research, clinical development, manufacturing, quality assurance, regulatory affairs, supply chains, investment, and due diligence. LSVP is currently available to teams and institutions, while Anthropic is working to expand it later to the individual Pro and Max plans.
Two access tracks based on risk level
Joining requires verification of the applicant’s research qualifications, safety standards, and ethical oversight mechanisms. After passing verification, teams can apply for one of two types of grants.
The Standard Use track provides dedicated access for most life sciences work and covers entire teams and varied daily workloads, with annual renewal. It currently includes Mythos 5.1, Opus 5, and Sonnet 5, as well as future models when they launch. The High-risk Use track is a separate option for work that remains blocked under the standard track and removes all controls that block life sciences requests. However, it is tied to a single research project and requires renewal every six months.
Anthropic cites as an example of high-risk use a project studying how human immune pathways recognize a specific family of viral vectors. High-risk Use grants are available for the Claude Opus 5 and Claude Sonnet 5 models, while the company is working with the U.S. government to expand availability to the Claude Mythos model, which will remain limited at launch to a small number of organizations subject to additional verification.
From immediate blocking to pattern monitoring
The program’s design is based on what Anthropic calls “shared responsibility.” Instead of blocking every request that could potentially be harmful at the time it is submitted, the company monitors usage activity externally to look for patterns that could indicate misuse distributed across multiple sessions and requests. The threats the controls are designed to address include account compromise, insider threats, and misuse of software agents, particularly when they operate in swarms or as part of long-running tasks.
Each organization’s access is tied to the use cases it identifies in its grant application. If Anthropic detects activity outside the declared scope, it can notify the organization’s officials so they can take triage and remediation measures within agreed time frames. The company requires data associated with observed activity to be retained for 30 days. It says this data is isolated, is not used to train models, and cannot be accessed by Anthropic’s life sciences research teams. Other cybersecurity controls remain active under LSVP grants.
Availability and current limitations
The program is available in Anthropic’s Console for API use, as well as through the Claude for Enterprise and Team plans, and users can switch between grants within the API and Claude Science. In Claude.ai and Claude Code, a predefined default grant is initially applied, except when Claude Code is used with API authentication. The program does not currently support individual plans or third-party platforms.
LSVP is also unavailable during the pilot phase to organizations that use a BAA. Anthropic requires organizations handling protected health information (PHI) to use separate organizations without the BAA option enabled and that are not HIPAA-compliant.
Editorial reading: Expanding access while shifting part of the oversight
The actual change here is not the launch of a new model, but the creation of an institutional access layer that separates documented daily use from higher-risk projects. This could reduce friction for researchers, but it shifts part of enforcement from an automated decision on every request to subsequent monitoring that depends on the user organization and Anthropic’s ability to detect unauthorized patterns. Retaining data for 30 days, along with excluding organizations with BAA enabled, also imposes important practical limitations on entities handling sensitive health information. The effectiveness of the new shared-responsibility model and its ability to prevent misuse remain open questions that the pilot program will need to answer.