A model cascade is a cost-optimization routing pattern where requests are tried against a cheap, fast model by default, and only escalated to a frontier model when a confidence check, complexity classifier, or the cheap model's own refusal signals that it can't handle the task. This differs from a flat fallback chain, which exists purely for reliability (retry on error or rate limit) rather than cost: a cascade is a deliberate quality-cost tradeoff applied to every request, not just failed ones. Enterprise routers built on OpenRouter or similar gateways implement cascades by pairing routing rules (task type, prompt length, tool-call complexity) with a cost ceiling, so routine requests never touch a frontier-priced model.