Latent Space asked how Anthropic now decides to release a frontier model such as Claude Opus 5.5, and the answer started with policy. The conversation described preventing distillation, running safeguards both while a model works and afterward, and setting rules for each model as now making up most of the work of a launch. None of it, the discussion said, was work the team used to have to consider when launching models.
Latent Space published the conversation on October 27, 2026. The guests were Katelyn Lesse and Angela Jiang, who work on Anthropic's Claude Platform, the set of services developers use to build on Claude. During the conversation, Opus 5.5 was described as having launched the day before. Anthropic announced the model on September 22, 2026, so the recording likely took place about a month before the episode came out. The announcement describes external evaluation before release, behavioral auditing and safeguards built for that model.
How releases changed
The conversation opened with recent history. When Anthropic's Fable model first came out, it was recalled, there was Glasswing and a delay and change in how models were released, and safety testing became standard for most frontier models. Project Glasswing, announced April 7, 2026, gave selected infrastructure providers and software maintainers restricted access to a model called Claude Mythos Preview for defensive security work. Anthropic backed it with up to $100 million in usage credits.
Retention rules, set model by model
The discussion explained that some abuse can't be caught while a model is answering. It described large networks of accounts that people create to misuse a model or to distill it. Distillation means collecting a model's outputs to train a cheaper copy of its abilities. As text streams through the system one request at a time, there is no way to tell that something malicious is happening.
The response described has been data retention: keeping records so that automated systems and human reviewers can go back over the activity. According to the discussion, Anthropic now decides for each model whether to require it, and Opus 5.5 does not carry the retention requirements that Fable has. Whatever policy a model gets, the team then has to do the work of enforcing it.
The conversation also took up criticism that Fable offered no zero data retention, an arrangement in which the provider stores none of a customer's prompts or outputs. The counterpoint raised was that critics were forgetting the context: Fable had been pulled back, and distillation attacks are real, so careful handling of the problem was welcome.
Vetted access for dual-use work
Verification programs are another part of launch work. They change how the online safeguards apply for approved users. The discussion named Anthropic's cyber verification program and its life sciences verification program, which was described as new at the time.
Jiang explained the reasoning in terms of dual use, meaning abilities that can serve both good and harmful ends. Safety evaluation scales up with capability, she said, and the most attention goes to abilities that spike in particular areas, judged by whether they favor attackers or defenders. Coding is the easy case: "great, go make awesome products. Go ship more things." Biology and cybersecurity are harder. A legitimate security researcher hunting for vulnerabilities, or a scientist doing biology research, needs some of the same capabilities a bad actor would want.
Jiang walked through how access works. A defensive security user who wants advanced cyber capabilities on a model such as Opus applies with a stated use case. If the application is approved, Anthropic will "diminish the safeguards on your version of the model." That requires understanding each use case and limiting access to a particular user, she said, "because we wouldn't do that generally."
Anthropic's published material gives more detail. The Life Sciences Verification Program, announced September 17, 2026, offers two kinds of grant. Standard Use grants cover a whole team and are renewed every year. They relax the biology classifiers (automated filters that flag or block risky requests) for ordinary scientific work. High-risk Use grants are tied to a specific project and are renewed every six months. They remove life-sciences blocking for that project, but other safeguards stay on. Applicants are checked for research credentials, security and ethical oversight. Monitoring requires 30-day data retention, and that data is kept separate from model training.
The Cyber Verification Program expanded in October 2026 from one access level to three: Defense Access, Red Team Access and Specialized Access. Each covers more sensitive security work than the one before. Approval covers only the organization that applied. Reduced classifier blocking comes with verification and security controls, and retained data supports misuse monitoring. An October 6 update on the Glasswing page says existing Glasswing members would move into the Specialized Access tier.
The models themselves are split the same way. Anthropic's Fable 5.1 and Mythos 5.1 announcement describes one underlying model offered with different safeguards. Fable is for general access, and Mythos is available through trusted-access programs.
Trust grants
All of this depends on knowing exactly who is allowed to do what. The conversation described a system built to carry the authorization to use a model, or to join a program with relaxed safeguards, through the whole of Anthropic's infrastructure. Getting access to Mythos, it was explained, does not just mean being handed an API key (the password-like credential a program uses to call the model) while someone flips a feature flag, a switch that turns on a capability for an account.
The discussion compared these "trust grants" with the identity and access management (IAM) permissions on Amazon's cloud, and with more elaborate authorization designs such as Google's Zanzibar. The answer was: not much different. Ordinary authorization checks that a request touches only your own data. A trust grant is a clean record saying that someone has been given access to a specific model or a specific program. A service sits on the path of every request and checks whether the caller holds the grant for what it is trying to do. One example was given. When Fable 5.1's handling of preserved thinking changed, some existing customers got extra time to update their integrations, and each was tracked by issuing them a grant.
Jiang said the team had debated how generic to make this access system. A normal product instinct, she said, would be to focus on the specific use cases. The team chose the more general design "because we were, like, we don't know." The system then became "massively useful" for general access to models, for managing any of Anthropic's trusted-access programs and for handling de-deployments, when a model is withdrawn. Jiang tied this to the basic building blocks, or primitives, that Lesse had been describing. She said the generic approach "pays, like, increasing dividends."
One button to pull a model back
The conversation turned to a habit familiar from platform engineering: after one incident, a team overprepares for it, and the incident never happens again. The discussion described that kind of preparation at Anthropic: a system that lets it pull a model back for everybody with a single button.
According to the discussion, the team had debated whether the work was justified, since it hopes nothing like that ever happens again. What justified it was that the button extended primitives the team needed anyway for the verification programs and for access to models like Mythos. Putting the same controls under everyone's model access meant they would be in place if they were ever needed.
Planning for things nobody predicted
Jiang said decisions like this have changed how the team builds its roadmap. In traditional product management, she said, you guess what will happen and in roughly what order, then pick concrete bets. Then came the de-deployment of Fable, which she said the government drove: "I don't think any one of us would have guessed" that it would happen.
Yet the probability was clearly not zero, Jiang said. Product planning has become "deeply probabilistic." The team builds a portfolio of bets and prepares hedges for scenarios it does not expect to happen again. More often than not, she said, "we actually do use the thing." She was less sure about the rollback button itself: "we'll see about this one in particular."
What "pacing" means
The conversation then raised Anthropic's support for pacing, a slower and more careful rollout of more powerful AI. It noted that advocating pacing while releasing a new frontier model 11 days later can look contradictory on the surface, and suggested this simply meant the model had passed Anthropic's safety checks.
The answer described pacing in three layers. The first is the security of the architecture and infrastructure. The second is online safeguards that run while a model is doing its work. The third is offline classification of stored data. Opus 5.5 was described as performing very well on all three, which was given as the reason for feeling good about releasing it.
The safeguards still cost users something. When a classifier flags a request, the user can be moved to a different model partway through their work. The discussion acknowledged that this is frustrating and said work continues to reduce false positives.