A fresh "hello" came back badged with a model that has no function calling at all. Nothing looked broken — it answers well enough — but that turn, and every turn behind it, ran with no web search, no picture or video generation, no memory and no canvas. And when someone asked why it could not search, the model explained the gap as its own limitation.

That is the worst shape a routing defect can take: everything works, nothing errors, and the product quietly becomes less than it is.

Ranked on what it can actually do

Auto scores every eligible model and picks a winner. A model with no tool rating at all was being treated as unrated-but-not-banned, which put it above a genuinely weak tool-caller. Reasonable in the abstract, wrong in this case: it was not weak at calling tools, it could not call them.

Candidates are now ranked on what they can really invoke. One model that could not answer the first question a session asked is out of the pool entirely.

Telling the model who it is

The other half is stranger. Nothing in the system prompt ever told the model which engine was running it. So it introduced itself as whatever it had been trained to say, and readers concluded the product and that one engine were the same thing — then disliked that engine and left, never finding the picker.

Every turn now carries the engine and provider by name, how many models are available, whether Auto chose or you did, and where to change it. On the phone there is also a tappable chip above the message box, plus /model and /auto, which previously went to the engine as a paid message.

One key, one account

A key one customer added for their own provider was making that provider's whole catalogue look available to everyone. No key was ever read or sent by anyone but its owner — but Auto believed those models were funded, picked one, and the turn died for everybody else. Availability is now scoped to whoever is asking.