Anthropic Restricts Opus 5.5 from Assisting with LLM Development

Anthropic has added classifiers to Opus 5.5 — similar to those used in its Fable models — that detect requests related to low-level LLM development (e.g., code for AI accelerators) and automatically redirect them to a less capable model. A notification informs the user of the switch, which persists for the rest of the conversation unless manually reverted. The filter targets a narrow set of tasks; Anthropic says ordinary ML work and programming are unaffected. In the API, there is no automatic redirect — the request simply stops until the developer configures a fallback model. Anthropic frames the measure as safety that must scale alongside model capabilities.

Why Claude switched models in your conversation