Beating AI News Flash: Codex source code shows that the server can reroute requests to other models after safety checks. Recently, a Pro user discovered that although they selected GPT-6 Astra throughout and the interface consistently displayed Astra, a large number of requests in the Usage records were actually running on GPT-5.5, with no switch notification. This issue has not yet been confirmed by OpenAI, but has been tagged as `bug` and `safety-check`.
Anthropic has long had a similar mechanism. After Fable 5 triggers a safety check, it automatically downgrades to Opus 4.8, and the company said from day one that it would notify users. But Claude Code has since been repeatedly reported to actually perform silent switches, with the interface even still displaying Fable 5.
Model routing itself is not strange. What is truly annoying is that you pay to choose a model, the interface also says it is that model, but after the backend swaps it out, it may not tell you.

