Everyone is building LLM routers, we deprecated ours

Manifest launched an LLM router in March to reduce costs by routing requests to models based on complexity but deprecated it after mixed results and challenges in accurately assessing task complexity from prompts alone. They found caching more effective for cost reduction and emphasized the importance of engineers choosing models deliberately.