A different way of saying no
Anthropic opened applications on Thursday to a Life Sciences Verification Program that gives vetted researchers versions of Claude with the biology and chemistry safeguards loosened, in exchange for identity checks, institutional oversight and monitoring after the fact rather than refusals in the moment.
The company said it has already onboarded dozens of organisations through an early access programme and expects to enrol hundreds within the first week. Access is currently limited to a set of US organisations, and Anthropic says it is working to widen that.
The problem being solved is a familiar complaint in drug discovery. Safeguards tuned to keep a general-purpose assistant away from bioweapons questions also refuse a great deal of ordinary laboratory work, and a false refusal in a research workflow is not a minor annoyance — it is the reason a team stops using the tool.
Two kinds of grant
The programme has two tiers. A Standard Use grant covers Mythos 5.1, Opus 5 and Sonnet 5 for what Anthropic calls typical life science workflows: drug discovery, research biology, clinical development and manufacturing. Those grants are issued to teams and renew annually.
A High-risk Use grant is a project-specific add-on that removes the remaining life-sciences safeguards for work still blocked under the standard tier. Those renew every six months. Anthropic says the high-risk tier is currently available for Opus 5 and Sonnet 5, while access to Mythos — the most capable tier — remains limited pending coordination with the US government.
That last detail is the structural change. For the highest-capability biology work, the gatekeeper is no longer Anthropic’s commercial judgement alone.
Vetting instead of classifiers
Every applicant goes through a review of research credentials, security standards and ethical research oversight. Access is then tied to the use cases the applicant described, and Anthropic says it monitors for activity outside that stated scope.
The monitoring model is the part safety researchers will argue about. Instead of a classifier blocking a prompt as it arrives, flagged activity is reviewed offline, with a 30-day retention window for the data involved. Anthropic says that data is compartmentalised, is not used to train models and is not available to its own life sciences research team.
It is a trade, and Anthropic is explicit that it is one. Real-time refusal stops misuse at the door but stops a great deal of legitimate work with it. Verification plus after-the-fact review lets the legitimate work through and accepts that misuse would be caught later rather than prevented.
Where it runs
The programme is available through Claude Science, Claude.ai, Claude Code and the API on Enterprise and Team plans, with individual Pro and Max plans to follow. It sits alongside Anthropic’s existing science programmes, which have offered free or discounted access to researchers rather than different safeguards.
Two things are worth watching. The first is whether other labs follow: if verified access becomes the normal way to do frontier biology, the vetting process itself becomes safety-critical infrastructure, and nobody has yet published how well such a process holds up under pressure. The second is what the US government’s role turns into. Coordination on who may use Mythos for biology is a narrow arrangement today. It is also a precedent.
