Your request may not run on the model you chose: Anthropic's safety routing breaks a quiet contract

The reported fact is specific and, for anyone building on these APIs, consequential: The New Stack reports that a request explicitly addressed to Claude Sonnet 5.5 may be routed by Anthropic to Sonnet 5 in situations the company deems "higher-risk." We only have the report's framing to work from, so we'll confine the commentary to what that arrangement means rather than speculating about the mechanism behind it.

Start with the case in favor, because it's real. If a provider concludes that one model version handles a category of sensitive request more safely than another, silently steering that traffic to the safer model is a rational safety control — arguably more responsible than refusing outright and leaving the user with nothing. Safety systems that operate below the API surface are not new, and every major lab runs classifiers users never see.

The problem is the contract. A model identifier has, until now, functioned as a promise about what code runs: the thing developers pin in configuration, benchmark against, write evaluations for, and cite in compliance documentation. If that identifier becomes a preference the provider may override, then reproducibility degrades in ways that are hard to detect — your evaluation suite passes on one version while a subset of production traffic quietly runs on another, and the behavioral delta shows up as an unexplainable bug report.

There's a direct line from here to the benchmark-integrity problems we've been tracking. Published scores are attached to model names. If the name no longer deterministically maps to the weights serving a given request, then every third-party comparison carries an unstated asterisk — and the regulated industries that most need auditability are exactly the ones that route the most "higher-risk" prompts.

Our read: the fix isn't to abandon safety routing, it's to make it legible. A response header naming the model that actually served the request, documentation of which categories trigger a substitution, and an opt-out for customers whose compliance posture requires determinism would resolve nearly all of this at modest engineering cost. We expect the industry to land there, because enterprise buyers will demand it before regulators do. The broader shift is worth naming, though: we are moving from models-as-products to models-as-managed-services, where the provider actively mediates what you get. That's probably where the technology has to go to be deployed safely at scale — but it only works if the mediation is visible, and right now it isn't.