Codex Usage (2026): Which Number Is the Real One
Codex usage confused me for a fortnight: three surfaces, three answers, none of them wrong. These are notes on which page answers which question.
Build a starter site → Official site
Three surfaces report codex usage
They are not the same meter. Inside the CLI, /status prints the percentage used and the reset time that OpenAI reports for your account. The subscription view lives at chatgpt.com/codex/settings/usage. Separately, platform.openai.com/usage is the API platform's own dashboard, covering metered requests and spend for anyone paying per token through an API key rather than through a plan. Most of the confusion I see online comes from reading a figure off one of those and comparing it to a figure off another, then concluding something is broken. Nothing is broken. They answer different questions. Work out first whether you are asking about plan allowance or about a bill, and only then open the page that covers it.
Why /status deserves to be the default check
It is free, it takes a second, and it reports what the provider says rather than what you calculated. There is no browser tab, no account switcher, no waiting for a chart to load. For anyone in the middle of a long refactor, that matters more than presentation: you get a percentage and a reset time in the terminal you are already sitting in. The habit worth forming is checking before starting something long rather than after it stops. A reset time two hours away changes what you should hand to an agent right now, and knowing that in advance is the difference between splitting a job sensibly and watching it halt in a half-applied state.
The five-hour window is not the current story
OpenAI's Codex pricing documentation describes a five-hour window that local messages and cloud chats once shared, and records that the window was suspended on 12 July 2026. It adds that additional weekly limits may apply and that task characteristics affect consumption. Read those together and the practical conclusion is uncomfortable but clear: there is no fixed token budget to plan against, no universal monthly cap, and no per-token rolling return anyone can model in a spreadsheet. The documentation points readers back at the usage dashboard or /status for current limits, which is an honest admission that the authoritative value is the one reported to your account. Anything you inferred from last month's behaviour describes a policy that has already changed once.
Menu-bar trackers, and what they buy
Third-party tools exist because people got tired of typing the same command. SessionWatcher, an independent macOS tracker, keeps provider-reported percentages and reset times in the menu bar beside local token totals and an estimated API-equivalent cost; its listing states macOS 14 or later, a seven-day bundle trial with no card, and a Codex-only licence at 6.99 US dollars one-time. Worth being clear about what that does and does not change: it surfaces the same numbers OpenAI reports, more often and without a keystroke. It does not raise a limit, predict one, or give you a cap the provider has not published. Convenience is the product, and for some workflows convenience is genuinely worth a one-time fee.
Planning around a ceiling you cannot see
Since the exact limits vary by plan and by task, the only forecast with any accuracy is your own history. Note what a typical review, a typical refactor and a typical test-fixing run actually consumed over a week, and you will have a better predictor than any published figure. Keep long jobs splittable so an interruption costs a step rather than a session. Keep something cheap in reserve for the work that does not need an agent reasoning over a repository at all: generating a landing page, spinning up a starter project, cloning a layout. That last category is where a separate tool earns its place, and it is why Begin.sh sits in the comparison below.
Four things worth knowing first
Two dashboards, not one
The Codex settings usage page covers your plan. platform.openai.com/usage covers metered API spend. Opening the wrong one is the single most common reason people think their numbers vanished overnight.
The free answer is a command
A live read of percentage used and reset time, printed into the terminal you already have open. No login, no extra install, and it reflects what OpenAI reports for the account you are signed into.
Limits move
The shared five-hour window was suspended on 12 July 2026, and weekly limits may still apply. Any fixed budget you memorised is a snapshot of a policy, not a rule you can rely on.
Task shape matters
OpenAI's pricing documentation says task characteristics affect usage, so two prompts of similar length can cost very different amounts. Estimate from your own recent runs rather than from word counts.
Where each answer lives
| Where you look | Codex | Begin.sh |
|---|---|---|
| Live percentage remaining | /status in the CLI prints provider-reported percentage and reset time | No meter to watch; a build either returns a zip or it does not |
| Subscription view | chatgpt.com/codex/settings/usage | Not applicable |
| Metered API spend | platform.openai.com/usage, a separate dashboard behind a platform login | Not applicable |
| Published limits | Vary by plan and task; weekly limits may apply | Check the site for current terms |
| Five-hour window | Shared by local messages and cloud chats until suspended on 12 July 2026 | No rolling window to track |
| What comes out | Edits, reviews and tasks against a real repository | A downloadable static site or Expo app |
| What you have to run | A ChatGPT plan sign-in or an OpenAI API key | A browser; no hosting, backend or auth included |
Checking it in four moves
- Name the question
Decide whether you are asking about plan allowance or about metered API spend. Those live on different pages, and the two figures will never reconcile because they never measured the same thing. - Run /status first
Inside the CLI it returns the provider-reported percentage and reset time immediately, with no browser and no account switching. Do this before a long run rather than after one stops. - Confirm on the settings page
For the subscription view, open chatgpt.com/codex/settings/usage. Use it to confirm what the command already told you, not as an independent source that might disagree. - Build your own baseline
After a week of ordinary sessions, write down what a review, a refactor and a debugging run each consumed. Your own history forecasts the next run better than any published number.
Questions people actually ask
Where do I check codex usage right now?
Two official places. Run /status inside the CLI for a live provider-reported percentage and reset time, or open chatgpt.com/codex/settings/usage for the subscription view. Both report the same underlying account figures, so pick whichever is fewer clicks from where you already are.
Does platform.openai.com/usage show my plan usage?
No. That is the API platform's dashboard for metered API requests and spend, and it sits behind its own platform login. If you are using Codex through a ChatGPT plan rather than an API key, it will not describe the allowance that is limiting you.
Is there a fixed monthly token cap I can plan against?
OpenAI's Codex pricing documentation does not publish one. It says limits vary by plan and task, that weekly limits may apply, and it directs users to the usage dashboard or /status for current values. Treat any fixed figure you find elsewhere as someone's estimate.
Why did a short prompt cost more codex usage than a long one?
Because length is not the driver. The documentation states that task characteristics affect usage, and a short instruction that triggers a large repository exploration does far more work than a long prompt answered from context already loaded. Judge cost by what the task required.
Do I need a paid tracker to stay on top of this?
No. Trackers like SessionWatcher surface the same provider-reported numbers continuously instead of on demand, which some people find worth a one-time fee. None of them raises a limit or reveals a cap OpenAI has not published, so the free command remains the baseline.
Why is Begin.sh mentioned on a page about usage?
Because a share of what people spend an agent allowance on is scaffolding rather than engineering. Begin.sh turns a prompt or a URL into a static site or Expo app and hands back a zip. It is a narrow tool, it includes no hosting or auth, and it leaves your Codex allowance untouched.
Keep the scaffolding off your weekly allowance
Begin.sh turns a prompt, or a URL you want cloned, into a working static site or Expo app and gives you the zip. It runs on its own infrastructure, so putting up a throwaway page costs none of the limit you have been watching.
Build a starter site →