Failover ≠ quality routing
Cascade routers escalate hard prompts for quality. Failover only calls a secondary when the primary errors or times out. The traffic shape—and the bill—are different.
Outage windows dominate the math
At 99.9% success, failover is a rounding error. During a degraded hour or regional outage, a large share of jobs hit the secondary; model that window explicitly with request volume and success rates.
Failed primary attempts may still bill
Hard outages and 429s are often free. Timeouts after the prompt is accepted may bill input (or more). Set failure billing as a bound, then compare ideal primary vs failover overhead.
FAQ
Should failover use the same model family?
Not required. Many teams pair providers (e.g. OpenAI → Anthropic/Gemini) for independence. CentsPerToken prices whichever models you pick.