AI coding agents consume 10 to 50 times more tokens than standard chat prompts due to continuous planning, generating, and testing cycles. This intense consumption is creating visible financial pressure on enterprises. Uber reportedly exhausted its entire 2026 AI budget by April after expanding its use of Claude Code, and Gartner predicts that AI coding token costs will surpass average developer salaries by 2028.
While some companies respond by implementing usage limits, focusing solely on token consumption is considered counterproductive. Instead, tracking the cost per merged pull request provides a better measure of how much useful software an investment produces. Managing these costs requires routing tasks based on complexity rather than relying on premium models for every step.
Enterprises can optimize spending by reserving powerful frontier models for complex reasoning while directing routine code generation and testing to open-weight models hosted on-premises. A hybrid platform implemented at Hewlett Packard Enterprise (HPE) routed routine workloads from over 150 internal applications to on-premises open-weight models, resulting in a 30 to 50 percent drop in token costs.
Hosting capacity internally not only improves economics as usage scales but also keeps proprietary code secure within the enterprise. Data indicates that top power users can consume roughly 450 times more than light users on the same license. Therefore, rather than rationing access, companies need to adopt operating models that make heavy AI usage more financially sustainable.