Guide · Usage
How to Monitor OpenAI Codex Usage, Credits, and Token Costs
A Codex task launched through a ChatGPT entitlement or flexible credit balance is not the same billing stream as an application calling the OpenAI Platform API. The interfaces, permissions, and accounting units differ. Combining them into one spreadsheet without a billing-path column creates misleading totals.
Evidence key: Observed — the revision date the editorial source publishes for this article.
Source: ModelScale guide /articles/monitor-openai-codex-usage (reviewed 2026-09-17). Provider links in the reference should be checked before purchase.
1. Separate ChatGPT Codex credits from API billing
Label every record as subscription capacity, flexible agentic credits, or Platform API spend. If a team uses more than one path, assign owners and budgets separately before creating an overall AI total.
Official references Codex rate card ↗ OpenAI Usage API ↗
2. Read the Codex Usage panel
Open Codex settings and find the Usage panel or Usage Dashboard available to your plan and role. Depending on the account, it can show current limits, recent usage, remaining credits, purchasing options, or auto-reload controls. Treat the live panel as the source of truth because models, rates, and entitlements can change.
Check it after representative tasks rather than after every prompt. A repository-wide refactor, a small bug fix, and a long research task have very different profiles. Grouping them together hides the work that actually consumes the budget.
- Capture the task type, model, and date before starting a representative run.
- Review the Usage panel immediately after completion.
- Record whether the run succeeded without retries or manual rework.
- Repeat the same task class a few times before treating the average as a planning number.
Official references Codex rate card and Usage panel ↗ Flexible usage credits ↗
3. Watch the token mix, not only the total
Codex usage is token-based for most plans, but input, cached input, and output can consume credits at different rates. Large repository context raises input; repeated stable context may benefit from caching; verbose answers increase output. Tool calls and task complexity also affect the final amount.