How to track AI spend when you code with several tools
Five steps from a list of subscriptions to one estimated cost view and a monthly check against the real bill.
By ClaudeGPT Si Newsroom · Editorial team
· 2 min read
The quickest way to track AI spend as a developer is to pull every coding tool's usage into one estimated cost view, then check it against each provider's billing page once a month. UsageAtlas, a free and open-source desktop app, handles the first part; the steps below cover the rest.
1. List what you actually pay for
Write down every AI tool that touches your code and how it charges you. A typical list mixes billing styles: Claude Code on a plan or an API key, Codex through a ChatGPT plan, Cursor with included usage, and GitHub Copilot measured in AI credits. Knowing the unit matters, because a plan window shows percentages, a credit system shows credits, and an API key shows tokens and dollars.
2. Check each tool's own meter once
Learn where each one reports usage, so you can confirm numbers later.
| Tool | Built-in check |
|---|---|
| Claude Code | /usage in the CLI; the dollar figure is an estimate at list price |
| Codex | /status in the CLI and the Codex usage dashboard |
| Cursor | The Cursor dashboard, where spend limits also live |
| GitHub Copilot | Metered usage in GitHub billing settings |
These are fine for a quick look. The trouble is that each shows only itself.
3. Put everything on one dashboard
This is where UsageAtlas comes in. It runs on Windows, macOS and Linux and reads supported local usage history and provider data. Codex, Claude Code, Cursor, OpenCode, Antigravity CLI and Pi / OMP can add estimated cost, split by day, tool and model. Copilot, Warp, Kimi Code and a few other credit-based tools show their reported allowances and balances.
A few things make it suited to spend tracking:
- The token mix sits beside the total, so you can tell whether cache reads, output or a pricier model caused a spike.
- Models with unknown prices show as unavailable, not as free.
- No account is needed, and prompts and responses are not read.
Keep in mind that every cost figure here is an estimate from tokens and public model prices.
4. Set a baseline and an alert
After a normal week, note your typical daily estimate. Anything well above it deserves a look. UsageAtlas can also send one desktop alert per usage window when you pass a threshold you choose, which helps you avoid hitting a plan limit mid-task. For hard spending caps, use the provider controls: Cursor spend limits, a Copilot budget, or workspace spend limits in the Claude Console.
5. Reconcile with the real bill
Once a month, put the tracker's estimate next to each provider's invoice or billing page. They will not match to the cent, since plans, discounts and contracted rates differ from list prices. The point is to catch drift: a tool that costs far more than you thought, or a model you forgot you switched to.
For a team that runs everything through one API account, that account's usage page and its per-user reports may be enough. For an individual juggling three or four assistants, the one-dashboard step saves the most time.
FAQ
Can I track AI spend without sharing my code? Yes. UsageAtlas reads usage data on your machine and does not read prompts or responses.
Does the tracker replace my billing page? No. It gives estimates for planning; the billing page is the record of what you owe.
What does UsageAtlas cost? It is free and open source.