The industry moved from seat-based licensing to consumption pricing and bills started drifting. Gartner has specifically criticised vendors for a lack of transparency here, and teams report jumps from twenty or a hundred dollars to thousands per developer per month.
Here is where that spend comes from and what actually helps.
Why the bill grows quietly
It comes down to how agents work. A chat request is one call. An agent working a task makes dozens, and each one drags everything that came before into context.
That gives three quiet sources of spend:
- Context accumulates. Step twenty costs more than step one because it carries the history of all previous steps.
- Retries. The agent errs, reads the error, tries again. Every iteration is billed.
- Background runs. An agent in CI fires on every commit, so spend tracks your team's activity rather than your decisions.
The most annoying line item
A figure that circulated widely across the industry puts a substantial share of tokens into fixing bugs the model itself generated.
That is worth sitting with: you pay twice, first to generate the problem and then to find and remove it. Which gives the practical takeaway that saves more than any micro-optimisation: the sharper the task going in, the less you pay coming out.
What actually lowers the bill
Split models by task. The flagship is for work that ties a lot of context together. Parsing logs, rewriting text and generating boilerplate are fine on a smaller model.
Cap answer length. Output costs several times input. Asking for "code only, no explanation" cuts the bill noticeably and breaks nothing.
Do not put a whole file in context when you need one method. Input is cheaper than output, not free, and on agentic runs it compounds.
Set the ceiling up front. A monthly cap with an alert at seventy percent solves the problem before it becomes unpleasant. In the Claudexia dashboard the cap and its thresholds are configurable per user and per team.
Seeing where the money goes
A total is useless until you can break it down. You need to know which key, which project and which model produced the spend.
We report usage per key, and sub-organisations let you separate teams and projects onto their own budgets. Nothing to reconcile by hand at month end.
In short
Agent spend is not a fixed line, it is a function of how you frame tasks and which models you pick. A ceiling, per-key breakdown and sane model selection remove most of the problem.