# Measure AI coding by software shipped, not tokens spent

> As AI agents consume vastly more tokens per task, companies should track cost per merged pull request instead of rationing usage.
- Title: Measure AI coding by software shipped, not tokens spent
- Summary: As AI agents consume vastly more tokens per task, companies should track cost per merged pull request instead of rationing usage. Companies should measure AI…
- Keywords: ai coding, cost measurement, ai agents, token pricing, software development, infrastructure, open-weight models, hpe, technology, science, Measure, coding
- Source: MIT Technology Review — https://technologyreview.com/2026/11/02/1147702/stop-counting-tokens-start-counting-what-ships
- Read time: 2 min
- Topics: ai coding, cost measurement, ai agents, token pricing, software development, infrastructure, open-weight models, hpe, technology, science
## AI agents burn through tokens fast
When developers give AI agents coding tasks, the agents plan, generate code, run tests, and repeat to fix errors. A single task uses ten to fifty times more tokens than a basic chat prompt.
## Token costs are already straining budgets
Uber reportedly exhausted its entire twenty twenty-six AI budget by April after expanding its use of Claude Code. GitHub Copilot moved from flat-rate pricing to metered credits.
## Costs could exceed developer salaries
Gartner predicts that AI coding token costs will surpass average developer salaries by twenty twenty-eight, creating pressure to limit expensive agent workflows.
## Rationing usage has a hidden cost
Blanket limits on usage can restrict workflows that deliver value. Cost per token measures consumption, not results. Some spending reflects waste, but some represents the work of shipping software.
## Track cost per merged pull request
A better measure is cost per merged pull request: total AI-assisted development cost divided by the number of changes that pass review and are merged. Track rework and defects alongside it.
## Route work to the right model
Agentic coding does not require the most powerful model at every step. Use frontier models for difficult planning and novel reasoning. Route code generation, test creation, and refactoring to open-weight models you operate yourself.
## Own capacity to lower per-call costs
With rented inference, every call adds to the bill. With owned capacity, most cost is committed upfront, so each additional call gets cheaper as usage grows. Ownership also keeps proprietary code inside the enterprise.
## HPE's hybrid model cuts token costs
At HPE, a hybrid AI platform routes routine workloads from more than one hundred fifty internal applications to open-weight models running on premises. The result: token costs down thirty to fifty percent.
## Power users drive most spending
Data shows that usage, not licensed headcount, drives spend. The top five percent of users generated most consumption, and the heaviest used roughly four hundred fifty times more than the lightest on the same license.
## Track two metrics before and after
HPE recommends tracking cost per active developer, because averaging across all licenses hides power users driving spend. Also track cost per merged pull request to connect AI spending to software that ships.
## Build for adoption, not rationing
Most enterprises will run hybrid: frontier APIs for demanding work, owned capacity for repeatable volume. Companies that lead AI-assisted development will build operating models where more AI use improves economics instead of undermining them.
## Key takeaway

Companies should measure AI coding value by software shipped, not tokens consumed, and use hybrid infrastructure to lower costs while enabling adoption.