How Databricks and Peers Keep AI Coding Bills From Outrunning the Gains

Agentic coding tools have measurably boosted developer velocity at Databricks, but like most large-scale adopters the company hit a wall: spend that grows exponentially and, unchecked, threatens to erase the efficiency it buys. Drawing on its own experience plus conversations with Stripe, Coinbase, Uber, and Ramp, Databricks lays out a playbook for a ‘dual mandate’ — give employees broad, low-friction access to AI tooling while holding aggregate cost to a roughly fixed envelope per user.

The biggest lever is chasing the ‘efficiency frontier’ rather than the intelligence frontier — the cheapest model that clears the quality bar for everyday coding, a target that moves faster than peak capability as near-weekly releases improve intelligence-per-dollar. Because public benchmarks poorly predict real coding performance, teams build internal evals to decide what actually beats their incumbents; Databricks rolled out GLM on that basis, while Stripe declined Opus 4.7 after it added cost without quality over 4.6. Capturing those switches requires escaping harness lock-in, which Databricks addresses with a meta-harness (its open-sourced Omnigent) that presents one interface while dispatching to different underlying harnesses and models.

Beyond model choice, the post details automated routing in three flavors — request-level proxies, task-level dispatch by complexity, and escalation/delegation pairings of cheap and expensive models (e.g., Claude’s Advisor Tool, Cognition’s Devin Fusion) — with Databricks’ Unity AI Gateway Smart Router cutting average task cost by more than 30%. Rounding out the toolkit are developer-facing visibility, tripwires and budgets, and token-overhead reduction, all tied together by an AI Gateway design pattern. The through-line: cost control at scale is now an infrastructure discipline, and Databricks has open-sourced its key pieces.