A new MIT Technology Review analysis argues that the industry’s obsession with AI token costs is measuring the wrong thing. When developers hand a coding task to an AI agent, it plans, writes code, runs tests and repeats the loop to fix errors — a single task can burn 10 to 50 times more tokens than a basic chat prompt. And the most useful agents rack up the most calls.
AI agent token costs (Image: MIT Technology Review)
Why Rationing AI Coding Tokens Backfires
The financial pressure is already visible. Uber reportedly exhausted its entire 2026 AI budget by April after expanding its use of Claude Code. GitHub Copilot moved from flat-rate pricing to metered credits, and Gartner predicts AI coding token costs will surpass average developer salaries by 2028.
The natural reaction — usage caps, seat limits, approvals before developers launch agents — can backfire by throttling the workflows delivering the most value. Cost per token measures consumption, not results. Some spending is pure waste like retries and runaway loops, but some is the actual work of shipping software, and a meter cannot distinguish them.
Cost Per Merged Pull Request: A Better AI Metric
The piece proposes cost per merged pull request: total AI-assisted development cost divided by the number of changes that pass review and get merged. Track rework and defects alongside it so cheaper changes don’t conceal lower quality. If that cost falls as AI usage rises without quality dropping, developers are converting consumption into finished software.
Getting there takes two moves: reserve expensive inference for hard planning and novel reasoning while routing routine code generation to open-weight models run on-premises, and own the capacity so each additional call gets cheaper. HPE‘s hybrid platform, routing routine work from 150+ internal applications to on-prem open-weight models, reportedly cut token costs 30 to 50% — and kept proprietary code inside the enterprise.
Also Read: The Agentic Web Is Here: MIT Technology Review Says AI Agents Are Quietly Reshaping the Internet and 9 Out of 10 AI Models Raised Their Own Crypto Spending Limits, Cornell and Yale Researchers Find
Source: MIT Technology Review — read the full story there.
