Databricks Says Model Routing Cuts Average AI Coding Task Cost by More Than 30 Percent

Databricks published internal data on the cost of running coding agents at scale, and the headline number circulating from it deserves separation into parts. In its own write-up, the company says its Smart Router, a stateful proxy that sits between the client and the underlying models and sends each request to the cheapest model able to answer it, cut average task cost by more than 30 percent. Output quality stayed roughly level compared with always calling the most expensive model available.

The larger figure, savings of up to 90 percent, comes from stacking routing with other measures: substituting open-weight models for frontier ones on easy work, trimming prompt and context overhead, and cutting redundant tool calls. That is a portfolio result at particular customers, not a property of the router itself.

Routing has become a standard layer rather than a differentiator. Fortune reported that Cursor, OpenRouter and Databricks all now ship routing products, each betting that a large share of production traffic does not need a frontier model at all. The economic logic is straightforward. If a meaningful fraction of agent calls are classification, formatting or short edits, paying frontier per-token rates for them is pure waste.

The caveat is measurement. "Quality roughly steady" is a judgment made by the same party selling the router, and quality regressions from downgraded routing tend to appear in long-horizon agent tasks rather than in single-turn evaluations.