Why Routing Every Coding Session to One Model Is Becoming a Costly Mistake

#ai #llm #machinelearning #programming

Most coding agents don't spend the majority of their time doing hard reasoning.

After looking through hundreds of coding agent traces, we found that frontier models like Claude Opus, GPT-5.5, and GLM-5.2 were often handling routine work such as searching files, rerunning tests, updating imports, and reading logs. Those steps are necessary, but they rarely need the strongest model available.

The problem is that most agents pick one model at the beginning of a session and stick with it until the task is finished.

Whether the agent is designing a new feature or fixing a missing import, every turn goes through the same expensive model.

Instead, we explored routing at the turn level rather than the session level. Simpler steps stay on cheaper models, while difficult reasoning automatically escalates to stronger ones before dropping back down when the work becomes routine again.

Here's an example of how a single turn is classified before selecting the model tier.

This shifts the optimization goal from cost per token to cost per solved task. In real coding workflows, not every interaction deserves frontier-level reasoning, so treating every turn the same ends up wasting both latency and money.

Our benchmarks showed that per-turn routing can maintain frontier-level coding performance while significantly reducing provider costs.

If you're interested in the routing strategy, benchmarks, execution traces, and implementation details, we published the full engineering write-up here: