When your model pick is only a suggestion: Anthropic reportedly reroutes Sonnet 5.5 to Sonnet 5

🕒 Published on Zendoric: September 30, 2026 · 00:34
✨ AI-generated · how it's made
The New Stack reports that Anthropic may serve a request with Claude Sonnet 5 even when the user explicitly selected Sonnet 5.5, in what it describes as "higher-risk" situations. The detail is small; the precedent is not. Model routing is quietly becoming a safety control — and that changes what a model selector actually promises.
The reported fact is narrow and we will not stretch it: according to The New Stack, a user who selects Claude Sonnet 5.5 may have that request handled by Sonnet 5 instead when Anthropic classifies the situation as "higher-risk." We only have the report's central claim here — the full piece did not reach us intact — so the trigger criteria, whether users are notified, and whether this touches the API, the consumer app or both remain unknown to us. We flag that gap rather than paper over it.
Our thesis: the model picker has always behaved like a contract. You name a version, you expect that version — because version identity is what makes outputs reproducible, evaluations meaningful and regression testing possible. Routing breaks that contract by design. For a casual chat, nobody notices. For a team that benchmarked a prompt against a specific checkpoint, or a regulated workflow that documents which system produced an answer, silent substitution is an audit problem, not a UX quirk.
There is a defensible reading too. Routing a sensitive request to a different checkpoint is a graduated response — softer than a flat refusal, which is the blunt instrument labs have leaned on for years. If a provider has evidence that one model handles a class of sensitive prompts better, sending the request there is arguably the responsible move. The problem is not the mechanism. It is the disclosure.
Our read: substitution that is documented, logged and visible is a legitimate safety tool; substitution that is silent is a trust tax the whole industry ends up paying. The fix is unglamorous and entirely achievable — a response field that states which model actually answered, and published criteria for when rerouting kicks in. Long term we want labs to have fine-grained safety controls instead of crude blocks; that is how AI graduates from a toy that says no into infrastructure you can build a hospital workflow on. But infrastructure earns trust through verifiability, not assurances. Until the routing rules are public, treat the version number in your dropdown as a preference, not a guarantee.
🔗 Related on Zendoric
- Your request may not run on the model you chose: Anthropic's safety routing breaks a quiet contract · 2026-09-29
- Anthropic's $40B compute deal makes Musk its landlord — and, reportedly, a check on its own safety · 2026-09-06
- Claude Haiku 5.5: Anthropic launches its fastest, cheapest small model for high-volume tasks · 2026-10-09


