Suppose you ask Claude for help with biology homework and mention "bio" once. A safety classifier might decide the request looks risky and hand it to a different model. That example came up on an episode of Latent Space published October 27, 2026. Katelyn Lesse and Angela Jiang appeared on the episode to explain the engineering behind Anthropic's Claude Platform.
Jiang's answer went beyond safety. She argued that safety fallbacks share a weakness with "routers," the systems that send easy jobs to cheap models and hard jobs to expensive ones: when they choose, they don't know enough about the task. She would prefer one consistent Claude that adjusts how hard it works as a task goes on, at least within Anthropic's own family of models.
What a fallback is
Anthropic runs checks on requests while Claude answers them. In an earlier Latent Space episode, Thariq Shihipar of Anthropic's Claude Code team described how this works. Probes read the model's internal signals, then a classifier runs, and the request can be handed to another model. That handoff is the fallback. A classifier in this sense is a separate system trained to flag particular kinds of content. The October conversation looked at the other side of the process: what the handoff is like for people using Claude, and what developers can do about it.
Anthropic's documentation on refusals and fallback describes the mechanics. When a classifier blocks a request, the API returns a refusal stop reason and, where one is available, a policy category. The block can come before the model writes anything or after part of the answer has already streamed out. In the second case, the partial output has to be treated as incomplete.
False positives and a dumb router
The conversation turned to the release of Opus 5.5, which was described as performing well enough on the dimensions that matter across Anthropic's safeguards that releasing it felt right. The approach, as explained in the discussion, was to aim for the best user experience possible under those constraints while continuing to reduce false positives, meaning harmless requests that get flagged by mistake. The discussion conceded that being switched to a different model in the middle of your work is "still a frustrating experience." It also acknowledged the safety case for accepting more false positives.
The conversation then compared fallbacks to the wider routing problem: should a smart model or a simple one decide where a request goes? The argument put to Jiang was that the detector behind a fallback is a separate system, "clearly not as smart" as the Fable model it guards, so it will sometimes misclassify. Each misclassification produces a fallback nobody needed, as in the homework example. Jiang said that once a request hits a classifier, set logic decides where it goes, and she called the routing comparison "very similar."
That raised a puzzle: how can a weaker model route a query it can't fully understand? The discussion turned to whether classifying is easier than answering, since a model might recognize that a query is dangerous even if it could do little with the query itself. Jiang offered what she called the "super simplifying" answer, which was to have the smarter model make the call. She said that comes with trade-offs and that she was not claiming it would be the right user experience.
Cut off five minutes in
One case described on the show showed how late a block can come. Fable was given its own safety card with a request to break down anything interesting in it. The model started working through the document. About five minutes in, the analysis reached cybersecurity, and the session rolled back to Opus. By then the model had done a lot of work and built up a lot of context. The switch came from the accumulated reasoning, "the whole chain," and the original question had not caused it.
What Jiang would rather see
Jiang described the experience she thinks would suit users better: a "more consistent unified experience with Claude." She contrasted it with the current one, in which a user falls back to another model and can then "double fall back." She said there are points in Anthropic's decision tree where fallbacks happen for safety reasons, but the result is "really convoluted" for the person on the other end.
She took the case of a good user, someone working through biology homework or assessing the cybersecurity section of a safety card. That is "totally reasonable," she said, yet the user gets blocked and moved to another model. In her ideal, Anthropic's models or classifiers would be smart enough to recognize that situation. They would answer the legitimate question concretely and leave out only the part that could enable harmful use of the capability.
What developers can choose today
Developers building on Anthropic's API already have options. According to the conversation, an app built on the API does not have to move a blocked conversation to a fallback model, even though Anthropic's own apps do that. Developers can turn on fallbacks that run on Anthropic's servers. They can manage fallbacks in their own code with helpers in Anthropic's software development kits (SDKs, the libraries developers use to call the API). Or they can rephrase the request and continue with the same model. The discussion also pointed to further experimentation with in-between approaches that would unblock people safely while keeping the experience good.
The documentation matches this account. Server-side fallback runs only when a developer turns it on, and it retries on a model suited to the policy category. Apps can also use SDK middleware or write their own retry rules. For some categories, Anthropic recommends no fallback at all.
The trouble with routers
Jiang said fallbacks raise the same user-experience problem as routers. Take a cost router, which picks a cheaper or pricier model depending on how complex a task looks. The idea makes sense, she said, but it is very hard to do well, because you don't know everything about a task at the start. Even if you understood every model perfectly, which nobody does, you still would not have perfect information about the task. So in her words there is "a lot of error" in any routing logic, "no matter how good a classifier you produce."
She said exposing the routing logic to customers would not fix this. A system might explain that a task to center a div (a basic layout job on a web page) goes to something super cheap, while a task to detect vulnerabilities goes to something really expensive. Showing all of that to the user would not be a great experience either, Jiang said.
Her alternative was dynamic effort: the system's sense of how complex a task is would change as it learns more about it. If that means switching among many different models, she said, the experience becomes "very, very disjointed." She suggested instead a single consistent model that is very cheap at its lowest effort level and can scale itself up continuously, keeping a consistent persona throughout. She said she meant this only within Anthropic's family of models. Across other providers' models, differences in personality make it "very, very complicated," and she said she did not know how to solve the problem at that level.
The conversation noted that Anthropic's own models were the territory she was responsible for. It closed on a further difficulty: working out what the user intends. People don't always say clearly what they want, and models don't always understand them. The example offered was asking a simple question and getting back a whole HTML landing page when a few lines of output would have done.