Codex usage limits: the 5-hour and weekly windows, by plan

Codex on ChatGPT Plus and Business is limited twice: a window that resets every five hours, and a weekly limit on top of it. ChatGPT Pro has no five-hour limit, only the week. OpenAI publishes estimated messages per five hours by model, not a fixed number, so the figure that counts is your own: /status in the CLI, or the usage dashboard.

The Codex 5-hour limit

On Plus and a standard Business seat, OpenAI's pricing page gives estimated local messages per five hours by model, read on 14 October 2026: 5–45 on GPT-6 Astra, 15–150 on GPT-6 Sol, 15–160 on GPT-6.1 Sol and 350–3,000 on GPT-6 Luna. In its words, "these estimates are not fixed message limits": the count depends on the model, the size of the task and whether it runs locally or in the cloud, and "cloud tasks may use more of your allowance than local messages". Local messages and cloud chats share one allowance.

The window was off for several weeks this summer. On 25 August 2026 OpenAI turned it back on for Plus; Thibault Sottiaux, who leads Codex engineering, said it smooths the load on OpenAI's compute and stops casual users spending the whole week in one sitting. He also said the five-hour limit stays off for Pro at $100 and $200 "for the upcoming months".

The Codex weekly limit

"Weekly limits may also apply," OpenAI's pricing page says of Plus and Business, and gives no figure. It says nothing about a week on Pro, but Codex records one there: on my machine every Pro reading carried a 10,080-minute window. Each five-hour window draws on the week, so once the week is spent a fresh window does not help: you wait for the weekly reset. Enterprise and Edu on flexible pricing "have no fixed rate limits. Usage scales with credits"; without it, they get the same per-seat limits as Plus.

How to check your Codex usage limit

• The usage dashboard, chatgpt.com/codex/settings/usage: current limits and reset times for the whole account.
• /status in a Codex CLI session: the model, approval policy and current token usage; OpenAI's pricing page points to it for your remaining limits.
• /usage: account token usage, with daily, weekly or cumulative, and a menu to "redeem an available earned reset".
• /statusline puts rate limits in the footer so they stay in view.

Each of these sees one account, and none keeps a history of the windows. A workspace's admin side, on Business and Enterprise: Codex analytics for a team.

What one machine's windows recorded

Codex writes its windows into every session file: each token_count record carries a rate_limits block with the window length in minutes, how much is used and when it resets. On 14 October 2026, 94 of the 220 transcripts in my ~/.codex carried one, across a move from Plus to Pro.

100%

the Plus five-hour window, 11 September

40%

the highest Plus weekly reading

100%

the Pro weekly window, 22 September

• Plus, 27 August to 18 September: 1,477 readings, each with a 300-minute and a 10,080-minute window. The five-hour window filled once; the week never passed 40%. On Plus the short window is the one you meet.
• Pro, 18 September to 13 October: 15,828 readings, each with one 10,080-minute window and no five-hour window at all. Codex records the plan as prolite. With nothing to slow a long day down, the week filled four days after the switch.

One heavy user on one machine is not a benchmark; it shows how the two shapes behave, not how many messages your plan gets.

What to do at the limit

• A turn already running "will be able to continue working on that turn, subject to fair use limits".
• Wait for the reset the dashboard shows, or redeem an earned reset from /usage if your account has one.
• Buy credits or upgrade: Pro removes the five-hour window, not the week.
• Switch the CLI to an API key: OpenAI then bills your Platform account per token at standard API rates.
• Carry on in another agent. In Caprock, Continue in… starts a new Claude Code, OpenCode or Gemini CLI session in the same folder, with a summary of the last one as its first message — which you read and edit before anything is sent — or a session on another of your Codex accounts.

Codex limits next to Claude's, in Caprock

Caprock is a free desktop app that records every Codex and Claude Code session on the machine, whether or not it started them. It reads Codex's windows from the transcripts in ~/.codex — no hook, nothing written into Codex's config — and shows them on the Plan limits panel under Codex's name, beside Claude's, with the time Codex wrote them. A window your plan does not have says "not reported by Codex" rather than showing a number.

• When one is near its limit, the panel says when it resets and how much the other agent has left: "Claude is near its limit… or switch to Codex — its weekly window is at 8%."
• Several Codex sign-ins sit side by side in Settings → Subscriptions, each with its own five-hour and weekly windows. Near 90% of a window, the New agent picker says so and offers an account with room; running sessions never move. Three per agent are free, your own sign-in included.
• The figure is as old as the last Codex session. Codex writes its windows into its transcripts, not to a live feed, so with no session running the reading can be hours old; the "as of" time says so.

The free, built-in answer is /status and the dashboard, and if you run one Codex account and nothing else, they are enough. Caprock is for when Codex is one of several agents or accounts, and you want every limit, and what each session cost at list price, in one record on your machine. Install Caprock.