The situation

Your teams use Claude Code, Codex and Copilot. The bill arrives per seat and per API key. One team doubles its usage in a month and nobody can say whether that was a hard feature, a runaway agent, or a model that was bigger than the job needed.

Capping keys stops the spend, and it stops the work with it. A cap on a key doesn't know which feature it is cutting off.

Three places a budget can sit

Budget onWhat it controlsWhat it can't tell you
The seatWho has accessWhat any of it was used on
The API keyHow much is consumedWhich piece of work it served
The piece of workWhat each item may use, and what it did useNothing it needs: this is the one to plan with

How to set it, step by step

  1. Estimate before approval. When work is planned, each item gets an estimate in tokens and in time, from how similar work went before. The person approving the plan sees the cost before saying yes.
  2. Roll up to the team and product. The team's budget for a sprint is the sum of its items. A product's budget is the sum of its teams. Nobody sets an abstract monthly number and hopes.
  3. Choose the model per step. Reading and summarising doesn't need the largest model. Writing a design or a change often does. Most overspend is one big model used for everything.
  4. Measure against the plan. Each item reports what it used. Over budget is a warning, not a failure: it says where the estimate was wrong, and the next estimate learns from it.
  5. Report per delivered item. The number a leader can use is tokens per shipped story, not tokens per month.

Five questions to ask your own numbers

  • Can you name the three features that used the most tokens last month?
  • How many tokens did one shipped story cost, on average?
  • Which work went over its estimate, and why?
  • How much of what your agents read came from cache, and how much was new?
  • Which steps used the biggest model when a smaller one would have done?

With Wakalix

Every plan line carries a budget in tokens and time, and the person approving the plan sees it first. Spend is shown by cycle, by agent and by skill, against what was planned, with what came from cache. Each skill has its own model, and steps such as plan checks and exit checks use no model at all. See Token budgets and the walkthrough Where the tokens go.