Databricks drove down AI coding spend 70%

The useful read on the piece is that this is not really a story about one clever router. It is a story about AI coding becoming an operations problem. Once hundreds or thousands of developers are using agents, spend can jump fast because pricing changes, model behavior shifts, harness updates alter token use, and usage expands faster than finance processes can react. Several people with direct experience said that this does happen in practice, especially now that teams have moved from flat subscription plans to token-priced enterprise usage. The comments also sharpened what “good” AI coding use looks like in 2026. The productive pattern is not “let the model code and walk away.” It is to separate planning from execution, use stronger models for design judgments, use cheaper ones for implementation and review loops, and keep humans focused on architecture, prioritization, and spotting when the model has quietly overbuilt the solution. That was the consensus among the pro-AI practitioners. The skeptical pushback was not that AI coding is useless. It was that unguided use turns greenfield projects into bloated brownfield systems frighteningly fast, and that claimed throughput gains are hard to verify without seeing shipping products, long-term maintenance costs, or proof that more code output translates into more business value. Where the conversation landed is pretty practical. Routing and cost controls look real, but they only work if you already treat coding agents like a managed production system with traces, benchmarks, and explicit workflow design. The bottleneck is shifting away from raw code generation and toward judgment, evaluation, and keeping agent-generated complexity from outrunning human understanding.