Codex Exposed for Silent Downgrades: Selecting GPT-6 Astra, Backend Might Actually Run GPT-5.5

Beating AI Newsflash, Codex source code reveals that the server can reroute requests to other models after safety checks. Recently, a Pro user discovered that despite selecting GPT-6 Astra throughout and the interface consistently displaying Astra, a significant number of requests in the Usage records were actually processed by GPT-5.5, with no notification of the switch. This issue has yet to be confirmed by OpenAI but has been labeled as a `bug` and `safety-check`. Anthropic has long had a similar mechanism. After triggering a safety check, Fable 5 automatically downgrades to Opus 4.8, and the company has stated since launch that users would be notified. However, Claude Code has been repeatedly exposed for silent switches, where the interface still displays Fable 5 even though the backend has changed. Model rerouting itself isn’t surprising. What truly frustrates users is paying for a specific model, seeing that model displayed on the interface, but not being informed when the backend switches it out. [Original Link]