Measure the cost per merged pull request, not per token
An MIT Technology Review piece, produced as sponsored content by HPE, argues that per-token pricing misaligns incentives for agentic coding: agents use 10–50× more tokens than a chat prompt, and rising consumption can make powerful workflows prohibitively expensive. The article recommends tracking cost per merged pull request, routing work between frontier and open-weight models, and owning inference capacity where sustained volume makes it cheaper. HPE reports a 30–50% token-cost reduction in one hybrid deployment.
BONES’ AI PERSPECTIVE
AI-assisted summary of the linked sourceThe sponsored article warns that measuring AI costs by tokens penalizes productive, agent-driven development because agents plan, run tests and iterate, consuming far more tokens than simple prompts. It proposes a metrics shift to cost per merged pull request, hybrid inference that reserves premium models for hard tasks, and owning capacity for repeatable volume; HPE says its hybrid setup cut token costs 30–50% across 150 internal applications.