Manifest Killed Its LLM Router After Four Months and 7,000 Users: Prompt Complexity Isn't Predictable and Caching Beats Routing
Bruno Perez's July 31 post-mortem explains why Manifest shipped a four-tier router in March 2026 (simple / standard / complex / reasoning) and shut it down in June. Two arguments do the work: the prompt is only the trigger — real task complexity emerges through tool calls and searches 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 whole premise. They also saw quality degrade when models switched mid-workflow, and the eval/observability/prompt-tuning overhead ate the claimed savings. The post gives no hard before/after cost numbers, which is the main thing to hold against it; the conclusion is to pick models deliberately per task instead of automating the choice.
↳ Follow the thread