A corporate push to maximize AI token use is giving way to model routing and tighter task selection as organizations scrutinize subscription and inference costs. AP reported that companies were pulling back from a springtime practice dubbed tokenmaxxing, which encouraged workers and systems to use as many AI tokens as possible. Moody's AI analytics head Vincent Gusdorf said token costs at some large organizations had been doubling nearly every other month.
A cited example put premium AI access at roughly $200 per developer per month, which becomes a multimillion-dollar annual expense across 20,000 developers. Organizations are adopting model routing so simple requests go to cheaper systems while difficult coding or research tasks use more capable premium models.
Open-weight Chinese models including Kimi and GLM were cited as lower-cost alternatives that approach some capabilities of leading U.S. hosted systems. The report did not identify a universal productivity measure or a single model that was cheapest and adequate for every task.
A token is a unit of text processed or generated by a model, and total cost depends on input, output, caching, tool calls and provider pricing. Premium models may justify higher cost for difficult tasks, while routing errors can send sensitive or complex work to an unsuitable system.
Productivity measurement must account for review time, rework, security controls and task completion rather than raw token consumption. As of the edition cutoff, The reported examples do not establish industry-wide cost growth or causal productivity effects across all workplaces. The available reports distinguish measurements and attributed claims from conclusions that had not yet been established. Where officials supplied numbers, those figures remain subject to revision as agencies reconcile records and publish later updates. Subsequent records may therefore change the totals without changing the initial chronology. The next factual records to examine are enterprise disclosures on AI spending and measured output and routing tools, smaller models and provider price changes.