You are told when a different model answered your round

When the model pinned on a thread has no usable credentials, the round is quietly moved onto another one so your work keeps running. That part is unchanged — but until now the record lied about it: the thread stored the model you asked for while a different model did the answering. A failing round then blamed a model that never ran, and the tokens were booked against it too.

Every round now records the model that actually answered. When that is not the model you picked, the response also carries the requested one alongside it, so the swap is visible on the thread itself — not only in the chat text. That matters for rounds nobody is watching: a delegation sub-thread, a generator agent, a scheduled run. Their result now carries the fact, and an operator sees a warning naming both models in the log.

Token usage and cost follow the same truth: they are booked against the model that served the round.

And when no model can serve a round at all — the pinned one has no credentials and nothing else in the catalogue does either — the round now fails, with a message in your language naming the situation and pointing at Settings → Language Models. Before, it ended as a success carrying a raw provider error, which anything reading the result had no way to tell apart from a real answer.