Matt Soellner

Today we launched Opus 5.5, which matches and even beats Mythos 5.1 on many bio-related tasks. We launched Opus 5.5 with the same safeguards as Fable 5.1, which means that most bio-related queries fall back to Opus 5. To use Opus 5.5 for life sciences work, enroll in our Life Science Verification Program (LSVP). I imagine that many in the bio community will be frustrated by this decision and I want to share more context on how I think about these things. I also want to say upfront—bio is one of our most significant areas of focus, and it’s why I’m here. When we set about designing the new LSVP, it was with the goal of providing broad access for life sciences professionals with far fewer annoying classifier misfires, while also protecting against many of the more sophisticated risks. Naively, it would seem that we have to choose between enabling broad access and protecting against misuse—I view this as a false choice. The key is to think of developing these safeguards as an engineering problem that requires the same level of investment, ingenuity, and ops execution as developing our frontier models, or any other large scale technical project. The risks we’re defending against—which, as we shared in our threat report a few weeks ago, are no longer hypothetical—were certainly not obvious to me until recently, and I imagine are also not top of mind for most biologists. When most people picture bio risk, they imagine a small group trying to develop a bioweapon in a single session, from their own account. The threats that we need to defend against are much more sophisticated than that. They include malicious work split over hundreds of distinct benign-looking sessions run from different accounts, and legitimate organizations’ accounts being hacked by a well-resourced actor or misused by a rogue employee. To protect against these risks and provide a better experience for our users who are doing legitimate life sciences work, in the LSVP we’ve introduced the concept of use-case grants. While it isn’t always possible to distinguish benign from harmful usage in isolation, it becomes much easier when comparing activity against a team’s own description of intended usage. You describe what your team works on at a high level, without revealing anything sensitive. Data are retained for 30 days, and automated offline review compares usage against that description and flags notable mismatches for a strictly compartmentalized review team. My team, the life sciences org, can never access this data. Because of these additional safeguards, for users in the LSVP we're able to tune down the classifiers so that they misfire far less often, without sacrificing on safety. I’ve described why I believe that the LSVP provides a better way for people to use our most capable models in bio, but one can still wonder: why are we implementing this now? There is no unambiguous threshold for when model capabilities reach a point that they start posing grave risks. But capabilities in bio, as in most other domains, progress very quickly and we should not be in the business of trying to time it perfectly. When in doubt, we’ll always choose to proceed with caution, and I’d much rather implement this new system too early than too late. The LSVP is currently in beta, which means that there are lots of improvements that we’re working on. We’re working to expand it so that individuals can get access—for example, for people developing or identifying treatments for a family member’s cancer or rare disease. Personally, I feel both the urgency to get our models into the hands of the broader life sciences community and a visceral sense of the stakes in getting it right. I think this is the right balance, and I look forward to seeing what people do with our most capable models. anthropic.com/news/life-scie…