Codex usage: how to check it, what counts, and when it resets

Short answer: Codex usage is governed by two limits at once — a rolling 5-hour window and a weekly cap — and both are set by your ChatGPT tier. The 5-hour window refills continuously; the weekly cap resets on a fixed 7-day schedule. What drains them is tokens: every message, response, file read, and log line adds up. You hit the wall fastest on token-heavy agent sessions. For the weekly ceiling specifically — when it resets and what happens when you hit it — see the Codex weekly limit guide.

How to check your Codex usage

There are two places to look, and they answer different questions.

The CLI is the quicker and more dependable one. Codex reports the state of your current session — how much of your allowance you’ve drawn down and roughly where you stand against the window — without depending on which version of the account UI you happen to be on.

Your ChatGPT account settings hold the plan-level view: which tier you’re on and what usage it has recorded. This is the “dashboard” people go looking for, and it’s a common source of confusion because it lives inside ChatGPT’s settings rather than on a separate Codex site, and OpenAI has moved and relabelled it more than once. If a bookmark stopped working, the view was most likely relocated rather than removed — check your account settings before concluding it’s gone.

Neither view tells you why your usage looks the way it does. For that you need a breakdown by category, which is what the audit further down produces.

What counts toward your limit

Codex doesn’t meter your time — it meters tokens. A short back-and-forth costs little; a long agent run that reads dozens of files, replays a failing test suite, and carries a growing conversation costs a lot. The context is re-processed on each turn, so noisy content you pulled in earlier keeps costing you until the session ends. That’s why two developers on the same plan can have wildly different mileage: one keeps sessions lean, the other floods the context with logs and re-reads.

The 5-hour rolling window

The first clock is a rolling five-hour window. Your first message starts a session, and everything you do for the next five hours draws from the same allowance. Heavy back-to-back work drains it quickly; when it empties, you wait for it to roll over. Because it’s rolling rather than fixed, it refills continuously — five hours after each session begins.

The weekly cap

The second clock is the weekly ceiling — a separate budget from the 5-hour window, reset on a fixed 7-day cycle, and the one that typically stops a heavy week even when each day stayed under the rolling window. When it is spent, waiting out the 5-hour clock does not help. The dedicated Codex weekly limit page covers that cap end to end: the two clocks, when it resets, and what happens when you hit it.

Limits by ChatGPT tier

Codex access rides on your ChatGPT subscription, and the allowances scale with it. Exact figures change as OpenAI tunes them, so treat the shape below as the pattern rather than a promise, and confirm against your current plan.

ChatGPT tier	5-hour window	Weekly cap
Plus	Standard allowance	Standard weekly ceiling
Pro	Substantially higher	Substantially higher
Business / Team	Per-seat, pooled options	Higher, seat-based
Enterprise	Negotiated / highest	Negotiated / highest

Relative scaling as of July 2026; check OpenAI’s current documentation for exact numbers, which change over time.

When it resets

The 5-hour window resets five hours after the session that started it — a moving target through the day. The weekly cap resets on a fixed cadence tied to your account’s 7-day cycle, not to Monday. If you’re capped, the app or CLI will tell you roughly when each clock frees up.

How to stretch your Codex limit

You can’t trim what you can’t measure. Codex writes a local record of your sessions, and that’s enough to break down exactly where your tokens went — tool output versus file re-reads versus conversation. The free hypermile-audit CLI reads those local records and prints a plain-English report: how much context you processed and which categories dominate. Nothing is uploaded; it runs on your machine and it’s open source.

Once you can see the breakdown, the fixes get concrete. Often a large share of a week is compressible log output, or a handful of files re-read many times over. Hypermile is built to trim exactly that noise before it reaches the model — and to forecast when you’ll hit your cap instead of surprising you at it.

Run the audit below to get your own numbers, then download Hypermile that trims the waste automatically. New to Claude’s version of these caps? See how Claude Code usage limits work and the Claude Code 5-hour limit, or compare the two products directly in Codex vs Claude Code.