Summary: Announced on September 17, 2026, Anthropic’s Life Sciences Verification Program gives verified teams more permissive access to Mythos, Opus, and Sonnet for covered biology work. The beta replaces some real-time blocking with offline behavior monitoring and requires 30-day retention for LSVP traffic. Its first phase supports API, Enterprise, and Team use, while excluding individual Pro and Max plans and third-party platforms.
Classifiers tuned for general availability decide which models anyone building on biology can actually reach. The Life Sciences Verification Program (LSVP), announced by Anthropic on September 17, 2026, tries to shift that boundary: verification buys a team Mythos, Opus, and Sonnet under safeguards that lean more permissive for science work, with everything else left as it was.
What a verified team actually gets
The program targets work the generally available Fable models will not take on today — drug discovery, research biology, clinical development, and manufacturing, per Anthropic’s announcement. Access is delivered through Claude Science, Claude.ai, Claude Code, and the API.
Planning comes down to two grant types. Standard Use spans most R&D workflows, can extend to an entire team, and renews yearly; today it applies to Mythos 5.1, Opus 5, and Sonnet 5, and it will cover future models as they ship. High-risk Use arrives as an add-on that clears the life-sciences blocks completely, but it binds to a single research project rather than a team and must be renewed every six months. High-risk grants for Opus 5 and Sonnet 5 are open now, the announcement notes, while Mythos high-risk access stays confined to a small set of entities pending work with the US government. Cyber classifiers and the remaining safeguards are unaffected.
Permission is only half of it; the other half is monitoring
Where enforcement sits is the engineering decision worth studying here. Real-time blocking rejects a request as it arrives; Anthropic says it is replacing that with offline monitoring that reads patterns of behavior instead, because serious misuse tends to scatter across many requests and sessions. Legitimate work gets interrupted less often, and the price shows up as data retention: LSVP traffic requires 30 days of retention for that monitoring to function. Per Anthropic, the data stays compartmentalized, never feeds model training, and remains off-limits to its own life sciences research teams.
That reframes the compliance conversation for a builder. The question is no longer simply whether a model will answer you; it is what telemetry your organization has to accept, and who inside your org is accountable when traffic drifts away from the use cases declared in the grant application. Anthropic calls this shared responsibility: it vets organizations on credentials and oversight, those organizations decide what safe usage means for their own people, and Anthropic routes out-of-scope patterns back to admins inside pre-agreed remediation timeframes. Access compromise, insider misuse, and agents taking unintended actions — swarms and long-horizon tasks included — are the threat models the announcement names.
Gaps in availability to plan around
LSVP is in beta and starts with teams and institutions. The first-party console covers API usage, and Claude for Enterprise and Team plans are supported. Not supported yet: individual Pro and Max plans, and third-party platforms. BAA-enabled orgs are excluded as well, which means customers handling PHI are pointed to separate non-BAA orgs without HIPAA coverage. In the API and Claude Science, grants can be switched natively; in Claude.ai and Claude Code, only a preselected default grant applies at the outset, with Claude Code on API authentication as the exception.
Details like that are what quietly break a demo. Say your workflow assumes a researcher can move between a standard grant and a project-specific high-risk grant from inside Claude Code — the supplied announcement says that portability does not exist yet.
How to read the timing
Dozens of organizations came in through early access, per Anthropic; hundreds are expected to enroll in the first week, and access for individuals is meant to widen over time. Pricing, application turnaround, and what the verification review actually inspects beyond research credentials, security standards, and ethical research oversight are all unspecified in the supplied announcement.
For anyone deciding whether to apply, the real question is narrower than “can I get access.” It is whether your team can state its intended use cases the way a job listing states duties — Anthropic says explicitly to leave out sensitive information and IP — and whether 30-day retention plus an admin who answers when flagged is something you can live with. The same institutional-trust calculus shows up for teams shipping public-interest science products; what Google’s science AI push changes about shipping public-interest products is a useful companion read. Write down the use cases and the incident owner before you touch the form.