The failover worked. The hour cost triple.

transcript

Your failover worked perfectly. And that hour cost three times as much.

Thursday, 14:10. Your model provider returns 529. Overloaded. For 40 minutes. So you move all traffic to a second provider. In one step.

On the primary, about 91 percent of input was read from cache. But the fallback has never seen your prompt prefix. Its cache is cold. Every input token is fresh. Full price.

Warm, a conversation costs about 2.2 cents. Cold, at similar prices, about 6.1. Roughly three times.

But here's the catch. It gets worse. Every one of those tokens counts against the fallback's input limit. Zero to full traffic triggers acceleration limits. Now the fallback answers 429 as well.

So shift in steps. 10 percent. Then 25. Then 50. Keep the prompt order cache-friendly on the fallback too. Pause optional features. And put a spend alert on the incident.

Failover has a price. Cost it before the bad evening. Then shift in steps.

Full breakdown in the lecture.