Manifest killed its LLM router after four months and 7,000 users.

Bruno Perez's post-mortem explains why they shipped a four-tier router in March and shut it down in June. Two arguments carry it: the prompt is only the trigger, since real task complexity emerges through tool calls after routing has already committed; and cache reads run 75-90% cheaper than uncached input, so the model stickiness that makes caching pay defeats the router's entire premise. Quality also degraded when models switched mid-workflow. No hard before/after cost numbers, which is the main thing to hold against it. Pick models deliberately per task instead of automating the choice.