Every AI provider has bad minutes. A model gets overloaded, a region hiccups, an endpoint answers with an error for ninety seconds. There is nothing unusual about that. What matters is what the app does during those ninety seconds.

Until this week, ours mostly gave up. Images, speech and transcription had been moving to a sibling model for months. Text never did, anywhere — so a provider having a bad minute ended an interactive answer exactly as it ended a workflow.

The same question, asked of another model

Every text call now goes through one path that knows how to ask elsewhere: the main chat turn and the tool steps behind it, Continue, deep research, the vision bridge, prompt improvement, the text step inside a workflow, and music renders. The replacement is not a random cheaper model. It is a sibling that has the capability the request actually needs.

One case is deliberately left out. Once a reply has started appearing on your screen, we leave it alone. Restarting on another model would mean rewriting or contradicting sentences you have already read, and that is worse than an honest error.

Paying the model that answered

The interesting half of this is the bill. Rerouting a call is easy; describing it honestly is not. Every reroute used to keep the first model's prices, so a call that failed over was billed at the rates of the model that had just failed. The chain now hands the caller the whole snapshot of whatever model really ran — prices, context window, markup — so the line in your billing history names the model that did the work, at that model's rate.

And when nothing works

Failing over is not always possible. When a run genuinely fails, the money for the steps that produced nothing now comes back to where it was taken from. That was the other half of the week, and it has its own post.