Quick answer
AI coding spend has two parts: fixed subscription or seat cost, and variable usage such as credits, on-demand requests, and API tokens. Track them in separate columns, because a rising bill is almost always the variable side. UsageAtlas shows reported plan usage and local token history, which is the variable input.
The two-column model
| Column | Examples | Behaviour |
|---|---|---|
| Fixed | Seats, plan price | Predictable monthly |
| Variable | Credits, on-demand, API | Follows usage |
| Indirect | Blocked time, retries | Hidden cost |
Most teams budget the fixed column and discover the variable one at invoice time. Reversing that order is the whole exercise.
Build the tracker
- List every tool and its billing type.
- Record the fixed monthly cost per tool.
- Record reported variable spend per period.
- Divide variable spend by completed tasks.
- Review monthly and note plan changes.
The cost per project guide covers attribution, and the team usage guide covers seat sizing.
The levers that move spend
- Model choice: the same task at a different tier.
- Context size: what gets resent on every turn.
- Attempts: retries are charged and rarely help.
- Seats: idle subscriptions are pure waste.
- Overflow settings: caps decide whether a heavy day becomes an invoice.
UsageAtlas makes the first three visible per day and per project. The budget guardrails guide turns them into policy.
Questions, answered
How do I track AI coding spend?
Separate fixed seat cost from variable usage, record both monthly, and divide variable spend by completed tasks so the number is comparable over time.
What is the biggest hidden cost?
Retries and long context. Both increase the charge without increasing output, and neither appears on a plan page.
Can a local tracker see spend?
It sees tokens and, where the provider reports them, credits and plan usage. The invoice remains the final authority.