Anthropic launches AI program for biology

Anthropic is opening access to its most capable AI models for laboratories, pharmaceutical companies and other life sciences teams, but only after verifying who will use them and why.

The company has introduced the Life Sciences Verification Program (LSVP), a beta program that lets biology professionals work with Mythos, Opus and Sonnet under more flexible rules than those applied to its models available to the general public. Anthropic has already brought dozens of organizations into an early access phase and is now accepting applications from more teams.

What the program allows

The LSVP is designed for tasks that Anthropic's general models often block because of their potential for misuse. These include:

• Drug discovery and development
• Biology and protein research
• Clinical development
• Manufacturing, quality control and regulatory affairs
• Investment analysis and biotechnology company evaluation

In practice, a laboratory could use Claude to analyze results, design experiments, review documentation or speed up a stage of drug development. Access will be available through Claude Science, Claude.ai, Claude Code and the API, depending on the account type and authorization.

Anthropic currently offers the Mythos 5.1, Opus 5 and Sonnet 5 models through the program. Access can be extended to an entire team and is renewed each year.

Two access levels

Before joining, each organization must demonstrate its research credentials, security standards and ethical oversight mechanisms. It can then apply for one of two types of authorization:

• Standard Use: covers most routine life sciences work and allows daily use by entire teams.
• High-risk Use: reserved for projects with greater misuse potential. It is granted for a specific project, not the entire team, and must be renewed every six months.

The second level removes the specific safeguards that block certain life sciences-related requests. For now, it is available for Claude Opus 5 and Claude Sonnet 5. High-risk access to Claude Mythos is limited to a small group of entities that undergo additional checks, while Anthropic works with the United States government to expand it.

Safety measures for other areas, such as cybersecurity, remain active.

Fewer blocks, more monitoring afterward

The most important change is how Anthropic will oversee usage. Instead of automatically rejecting every potentially dangerous request, the program will analyze behavior patterns after they occur. This allows legitimate research to move forward with fewer interruptions, but it also requires data related to activity flagged for review to be retained for 30 days.

Anthropic says that data will be isolated, will not be used to train models and cannot be accessed by its internal life sciences research teams. If it detects activity outside the authorized purpose, it may notify the organization's administrators, who will have to investigate and correct the problem within agreed timelines.

The system is designed to reduce three specific risks:

• Malware or account theft giving an attacker access
• An employee acting independently or being forced to use the account for malicious purposes
• An AI agent unexpectedly carrying out dangerous actions, especially during long or coordinated tasks

What changes for users

For now, the LSVP is available to organizations with Team and Enterprise plans, as well as through API access. It does not yet work with individual plans or third-party platforms. It is also unavailable to organizations with BAA agreements, meaning customers handling protected medical data will need to use separate organizations that are not HIPAA-covered.

For most authorized users, Standard Use permission will be sufficient. In Claude.ai and Claude Code, a default permission will be applied initially, while the API and Claude Science natively allow users to switch between authorizations.

Anthropic expects to bring in hundreds of organizations during the first week and expand the program in the following weeks. The central idea is clear: give scientific teams access to more useful tools without treating all biological uses as if they were the same. The challenge will be proving that post-use monitoring can detect abuse in time without turning legitimate research into an endless chain of reviews.