Databricks publishes measured cost controls for agentic coding at scale

Sources: primary, discussion
Summary: Databricks reports roughly 50 percent fewer generated tokens from harness and cache tuning, and over 30 percent lower average task cost from request routing. The post records negative evaluation results, including Stripe declining Opus 4.7 and Databricks measuring a cost regression from Opus 5.0 against 4.8.
Why it matters: Published negative model evaluations and named token-reduction techniques give teams a baseline to compare their own coding-agent spend against.