Stop counting tokens, start counting what ships

Coding agents don't answer once — they plan, write code, run tests, and loop to fix errors, which means a single task can consume 10 to 50 times the tokens of a basic chat prompt. The perverse result of per-token pricing: the workflows creating the most value are also the ones becoming most expensive, and the natural fix — usage caps, seat limits, approval gates — tends to throttle exactly the work that ships software.

The pressure is already concrete. Uber reportedly burned through its entire 2026 AI budget by April after scaling up Claude Code, and GitHub Copilot has shifted from flat-rate pricing to metered credits. Gartner projects that AI coding token costs will exceed average developer salaries by 2028.

The proposed alternative metric is cost per merged pull request: total AI-assisted development cost divided by changes that pass review and merge, tracked alongside rework and defects so cheaper changes can't hide lower quality. On that measure, falling cost per merge as AI usage rises — without quality slipping — means consumption is actually converting into shipped software.

Two levers improve the ratio. First, route by task: reserve frontier models (or on-prem flagships where policy demands) for hard planning and novel reasoning, while sending code generation, test creation, and repeated refactoring to self-operated open-weight models via a gateway that weighs complexity, data sensitivity, cost, and latency. The article cautions that benchmark proximity to frontier models doesn't establish equivalence across every repository — evaluate candidates against your own code, tests, and acceptance criteria. Second, change who owns capacity: rented inference adds cost with every call, while owned capacity front-loads spend so marginal calls get cheaper, with power, operations, maintenance, and depreciation all factored into a fully loaded cost per merged pull request.

Ownership also buys something absent from any invoice — control. With generation and testing on-premises, proprietary code stays inside the enterprise, and only approved planning prompts and governed exceptions leave through a controlled gateway, which opens sensitive repositories where security usually stalls adoption. HPE's hybrid platform offers a worked example: more than 150 internal applications route routine workloads to on-premises open-weight models, including on HPE Private Cloud AI, cutting token costs by 30 to 50%.

What to watch: whether cost per merged pull request becomes a standard engineering metric alongside DORA-style measures, whether gateway-based routing holds up on the sensitive repositories it's meant to unlock, and whether the HPE-style 30-50% token savings survive contact with fully loaded capacity costs across different organization sizes.