Databricks just cut its AI coding costs by 90%. Your token bill is next.

Databricks' engineering org was burning through AI tokens so fast that coding agent spend became one of its top three R&D line items.

A single runaway automation loop can burn a month's budget in an afternoon.

So they published their playbook. Four levers, stacked:

1. Route traffic to cheaper models. 50% savings. Most coding tasks don't need frontier intelligence.

2. Dynamic routing by task complexity. 30% savings. "Rename this component" doesn't need the same model as "design the architecture."

3. Visibility plus progressive friction instead of hard caps. 10% savings. Engineers who hit limits get a Slack nudge, not a shutdown.

4. Cut token overhead. 10% savings. Databricks found nearly half the tokens reaching models were context nobody asked for.

The result: per-task costs down 90% in some scenarios. Developer adoption went up, not down.

The lesson for every CIO watching their AI line item explode: the next cost crisis isn't GPU capacity. It's token waste.

Audit your AI coding agent spend this week. If you don't have per-engineer visibility into what tokens are actually producing, you're already behind.