Quick answer
Codex reports remaining capacity as a percentage for each window, not as a token balance. To turn that into tokens, divide your recent token consumption by the percentage it consumed. UsageAtlas shows the same percentage next to the token totals it reads from local session files.
Where the remaining number lives
| View | Shows | Refresh |
|---|---|---|
| Codex usage panel | Percentage left for the short and weekly windows | On open |
| `/status` in the TUI | Context usage and session tokens | Live |
| UsageAtlas Today | Percentage left plus reset countdown | Background |
| Provider billing page | Invoices and credit balance | Per billing cycle |
There is no public endpoint that returns an absolute token balance for a plan window. Anything claiming an exact number is estimating it from observed consumption.
Estimate the tokens behind the percentage
- Note the percentage before a work session.
- Note the tokens consumed during that session, from
/statusor a tracker. - Note the percentage after.
- Divide tokens by the percentage points consumed to get tokens per point.
- Multiply by the points remaining.
Example: 12,000 tokens moved the window from 80% to 76%, so one point is about 3,000 tokens and 76 points is roughly 228,000 tokens. Treat that as a working estimate, not a promise.
Watch the right number
The percentage is what the provider enforces, so plan against it. Tokens are useful for comparing tasks and models, not for predicting the exact request that will fail.
- Set an alert at 20% so you are not estimating under pressure.
- Track both windows; the short one can recover while the weekly one keeps falling.
- Re-estimate after a model change, because tokens per point shifts.
Remaining tokens are two different numbers
There is the context left in the current conversation and the allowance left in the usage window. They move at different speeds and neither is a prediction. Context drains with every file you add; the window drains with every model call, including the ones that only produce a short answer.
| Number | Resets when | Controlled by |
|---|---|---|
| Context remaining | You start a new session | Files and history in the thread |
| Session tokens | New session | Work done in this conversation |
| Window allowance | The window resets | All sessions across the account |
The usage limit guide covers what to do when the window number reaches zero.
Questions, answered
Why does Codex not show a token balance?
The allowance is defined in usage units rather than a fixed token count, so the provider computes it per plan and model. A percentage is the stable representation.
Is the remaining percentage per model?
No. It is per account window. Switching models changes how fast the percentage falls, not the window it draws from.
Does the percentage include cached input?
It reflects billable consumption, which counts cached input at the provider's cached rate. The percentage will not match a raw token sum exactly.