AI coding tools deliver immense value, but at scale, costs grow exponentially and threaten to overtake revenue. Databricks and other digital-native companies like Stripe, Coinbase, Uber, and Ramp have converged on a set of approaches to achieve a “dual mandate”: broad access to AI tooling with minimal friction, while keeping aggregate costs inside a fixed envelope per user. The post outlines three proven cost levers.

The biggest lever is moving to more efficient models as they are released. The “efficiency frontier” – models with the best intelligence per unit price – advances faster than the intelligence frontier. Companies like Databricks use automated evaluations, such as their benchmark that led to rolling out GLM internally. They also avoid costly model regressions, e.g., Stripe declined Opus 4.7 after finding no quality improvement over Opus 4.6.

A second lever is dynamic request and task routing. Databricks’ Unity AI Gateway Smart Router reduces average task cost by over 30% while matching quality of expensive models. Routing approaches include request-level routing, task-level meta-harnesses (like Omnigent), and escalation/delegation patterns. A third lever gives developers visibility, tripwires, and budgets to prevent unchecked spend.

Databricks open-sourced key infrastructure components – Omnigent and Unity AI Gateway – to help others replicate these gains. The techniques can be implemented with existing software or new infrastructure, enabling enterprises to push AI transformation without cost explosion.