Measuring coding agents by software shipped

AI coding agents plan, generate code, run tests, and retry, making a single task consume 10 to 50 times more tokens than a basic chat prompt. Per-token pricing can therefore make productive workflows expensive. Cost per merged pull request offers an alternative measure, connecting total AI-assisted development spending to changes that pass review and reach the codebase. Tracking rework and defects alongside spending helps guard against savings that compromise quality.

A hybrid approach reserves frontier models for difficult planning and novel reasoning while routing suitable routine work to self-operated open-weight models. A governed gateway can select models according to complexity, data sensitivity, cost, and latency. HPE reports that routing routine workloads from more than 150 internal applications to on-premises open-weight models reduced token costs by 30 to 50%. Models still need evaluation against an organization’s own code, tests, and acceptance criteria.

Owned infrastructure can improve economics under sustained demand and keep sensitive code inside the enterprise, but its full cost includes power, operations, maintenance, and depreciation. Low or unpredictable usage may favor APIs instead. HPE recommends establishing baseline costs per active developer and merged pull request, then testing owned capacity through a complete development cycle before deciding whether the operating model delivers better results.