Manifest Deprecated Their LLM Router and Explained Why

LLM routers are having a moment. The pattern is straightforward: intercept requests, score them by complexity or cost, then dispatch to whichever model fits. On paper, it optimizes spend and latency. In practice, the team at Manifest found the abstraction brittle enough to abandon.

The core problem with routers is that they insert a classification layer that must stay accurate as models change. Every time a new model releases — or an existing model gets a silent update — the routing logic drifts. What looked like a sensible heuristic at the time of calibration becomes a liability six weeks later. Maintaining that layer quietly becomes a second job.

There is also a composability issue. A router that works well for single-turn completions tends to break down when the call pattern becomes agentic or multi-step. The classification signal that routes a simple lookup correctly may misread a reasoning-heavy subtask. You end up either over-engineering the classifier or accepting that it will misfire on the cases that matter most.

What the deprecation signals for builders: the value proposition of a router assumes model capability tiers are stable and predictable. They are not. The gap between frontier and budget models narrows with each release cycle. When the performance delta shrinks, the marginal gain from smart routing shrinks with it, while the maintenance cost stays fixed.

The smarter path, based on what the team describes, appears to be direct model selection baked into the task definition rather than a runtime dispatch layer. Pick the model at design time, revisit when benchmarks shift, and skip the middleware.

For solo founders and small engineering teams, the lesson is operational: abstractions that require continuous calibration against an external moving target are expensive to own. Ship the simplest thing that works, watch the failure modes, and deprecate when the cost of ownership outweighs the benefit.

Source