What is AI token cost management?

Tokens are the units a language model bills on—roughly the pieces of text it reads and writes. Because you pay per token, and frontier models cost many times more per token than small ones, cost is a function of which model runs which task and how often, not of a seat count. That makes AI spend behave like cloud infrastructure or payment processing: metered, usage-driven, and capable of scaling faster than the value it produces.

It has been in the news for two connected reasons. Providers now sit at very different price points for broadly similar output on routine tasks, and routing infrastructure has become a business in its own right—Stripe's acquisition of OpenRouter put a price on the idea that token movement between providers is closer to a toll road than a tool. At the same time, some employers have started issuing per-employee token budgets alongside salary, which turns model spend into a line item managers are accountable for.

That collides with how budget cycles work. Software is usually approved annually as a fixed number, while token spend is variable and lands monthly with no cap unless someone sets one. Teams that forecast AI as a subscription tend to discover the overrun a quarter late, and the overrun rarely comes from one team—it comes from long context windows, retried calls, and multi-step agent runs where each step re-reads everything before it.

It affects more than engineering. Finance owns a forecast that now has a variable component. RevOps and marketing operations own the workflows that generate the calls—enrichment, summarization, routing, drafting—which is where most of the volume actually lives. Whoever signs the vendor contract owns the lock-in risk if a single provider carries the whole workflow.

The core discipline is matching the model to the job instead of defaulting everything to the most capable option. Classification, extraction, routing, and formatting are handled well by small, cheap models. Reasoning over messy context, drafting customer-facing copy, and multi-step planning are worth a frontier model. Batch and cache anything repetitive, trim the context you send, and keep at least one credible alternative provider wired up so price changes are negotiable.