Manifest has officially deprecated its LLM router, a feature originally launched in March as part of its LLM gateway to reduce inference costs. The router functioned by classifying incoming requests into four complexity tiers—simple, standard, complex, and reasoning—to select the most cost-effective model for each task. However, after four months of testing with 7,000 cloud users, the company concluded that the routing mechanism introduced more significant overhead than it saved. A primary technical failure was the inability to determine true task complexity from a prompt alone, as complexity often depends on external factors like tool calls or repository scale. Furthermore, the team found that prefix caching is a far more efficient cost-reduction strategy, as cache reads are 75% to 90% cheaper than uncached inputs; routing often disrupts the "stickiness" required for effective prefix caching. Additionally, the router introduced behavioral inconsistency and unpredictability, which complicates the management of automated agentic workflows and increases the burden on evaluations and observability. Ultimately, Manifest determined that the engineering effort required to manage the uncertainty of model switching outweighs the marginal savings, concluding that sticking to a single, battle-tested model is superior for most production use cases.