# Codex Usage Limits, Explained: The Five-Hour Window, the Weekly Cap, and Credits

> Codex draws from your ChatGPT plan: a five-hour window plus a weekly cap, shared with ChatGPT Work. What each plan includes, how resets and credits work, and your options at the limit.

Published October 16, 2026 · by the AQ team · Canonical: https://aq.dev/guides/codex-usage-limits-explained/

Codex runs on your ChatGPT subscription, and as of October 2026 that subscription carries up to two stacked usage limits: a five-hour window and a weekly window, and you need allowance remaining in both to keep working. The pool is shared, so Codex, ChatGPT Work, and (on plans where they are available) agentic features like ChatGPT for Excel all draw from the same allowance, whether a task runs locally in the CLI or as a cloud task. The big plan difference: ChatGPT Pro tiers ($100, $200, and $500 per month) currently have no five-hour limit in Work and Codex, while Plus and Business Standard seats do. OpenAI publishes estimated message ranges rather than token counts, and the fastest way to see your real remaining budget is /status inside a Codex CLI session or the usage dashboard at chatgpt.com/codex/settings/usage. This guide explains how each window works, what drains the pool, what every plan includes, and your options at the limit, verified on OpenAI's own documentation in October 2026.

## How does the Codex five-hour window work?

The five-hour limit sets how much usage is included in each window, and a new window starts when you send your first message in Work or Codex after the previous window ends (per OpenAI's help center, October 2026). It is a usage budget, not a time budget: you can reach the five-hour limit well before five hours have passed, even with weekly allowance remaining, because one heavy agentic run can spend the whole window's budget quickly. Light use may never touch it.

Not every plan has this window. As of October 2026, OpenAI's documentation says Pro 100, Pro 200, and Pro 500 currently have no five-hour usage limit in Work and Codex, and Business Premium seats drop it too. Plus and Business Standard seats keep it, with these estimated ranges for local messages per five-hour window (OpenAI's published estimates, October 2026, not guaranteed counts):

| Model | Local messages per 5-hour window (Plus and Business Standard) |
| --- | --- |
| GPT-6 Astra | 5 to 45 |
| GPT-6.1 Sol | 15 to 160 |
| GPT-6 Sol | 15 to 150 |
| GPT-6 Luna | 350 to 3,000 |

The spread inside each range is the task: a one-file edit sits at the top of the range, a long agentic run with lots of context, tool calls, and high reasoning effort sits at the bottom.

## How does the weekly limit work?

The weekly window sits above the five-hour sessions: a budget over a rolling weekly usage period, with the reset time shown in /status and in Settings then Usage. Hitting it stops Codex even when your current five-hour window has room. OpenAI does not publish weekly token figures for personal plans; the meter and its reset timestamp are the source of truth.

One mechanic is worth knowing before you ever buy a reset: a purchased instant reset (available to eligible Plus and Pro accounts) refreshes both the five-hour and weekly windows immediately, but it pulls your normal weekly allowance forward rather than adding extra usage, and it re-anchors your schedule. Your new weekly period starts with your first Work or Codex request after the reset, and the next automatic weekly reset lands seven days after that request. OpenAI's help center is explicit that a purchased reset applies immediately, cannot be saved for later, and generally is not refundable.

## What actually drains the shared pool?

Usage depends on the model, where the task runs, task complexity, context size, reasoning effort, speed mode, and tool use. The levers that matter most, per OpenAI's October 2026 documentation:

- **Model choice.** GPT-6 Astra consumes the allowance several times faster than GPT-6.1 Sol, and Luna is dramatically cheaper than both (visible in the message ranges above and in the credit rate card, where Astra output costs 100x Luna output per token). Switching models mid-thread does not restore allowance already spent; it only changes the burn rate going forward.
- **Reasoning effort.** Higher effort can use more allowance and, per OpenAI, does not always produce a better result. Their own guidance: Astra at low effort can outperform smaller models at high effort.
- **Speed modes.** Fast mode uses included allowance at 2.5x the standard rate. Ultrafast (GPT-6 Astra and GPT-6.1 Sol only) runs at 8x and is included only on Pro 500 and eligible Enterprise and Edu workspaces as of October 2026; buying credits on other plans does not enable it.
- **Where the task runs.** Local CLI messages and cloud tasks draw from the same pool, and OpenAI notes cloud tasks may use more allowance than local messages (a cloud run re-reads context and works longer unattended). Standard Codex Cloud environments add no separate VM charge; you pay in allowance, not rent. [Codex cloud tasks explained](https://aq.dev/guides/codex-cloud-tasks-explained/) covers that surface in depth.
- **What is free.** Auto-review safety checks (the reviewer that approves or denies eligible actions so Codex can continue) do not count toward usage limits when you are signed in with a ChatGPT account.

## What each plan includes, as of October 2026

| Plan | Price (October 2026) | Five-hour limit | Notes |
| --- | --- | --- | --- |
| Free / Go | $0 / $8 per month | Yes, small allowance | Codex included; no Codex Cloud |
| Plus | $20 per month | Yes | Weekly cap; credits and instant resets available |
| Pro | $100, $200, or $500 per month | Currently none | Included allowance still applies; Ultrafast on Pro 500 only |
| Business Standard seat | $25 per user per month ($20 billed annually) | Yes | Same message estimates as Plus |
| Business Premium seat | Per-seat upgrade | No | 5x the included usage of a Standard seat |
| Enterprise / Edu | Contact sales | Varies | Workspace credits or token-based USD billing with admin-set budgets |

All figures from OpenAI's pricing documentation and help center, October 2026. OpenAI does not publish absolute token allowances per plan, and the model lineup moves: GPT-5.4 left Codex for ChatGPT sign-ins on August 31, 2026, and GPT-5.5 followed on October 14, 2026, so a plan's practical capacity shifts with the models behind it.

## How do I check how much Codex usage I have left?

Three places, none of which consume usage:

```
# Inside a Codex CLI session
/status
```

/status reports your remaining five-hour and weekly allowance with reset times. The same meters live at chatgpt.com/codex/settings/usage and under Settings then Usage in the ChatGPT apps. On Enterprise workspaces with credit billing, members can also see per-chat credit usage in ChatGPT Desktop. If a limit lands mid-task, OpenAI's documented behavior is that Codex can finish the active turn, subject to fair-use limits, and then stops accepting new prompts until something changes.

## What are your options at the limit?

- **Wait.** A new five-hour window starts with your first message after the old window ends; the weekly reset time is shown in /status.
- **Switch to a cheaper model for the remaining allowance.** It does not refund anything, but the same weekly budget goes much further on GPT-6.1 Sol or Luna than on Astra.
- **Add credits.** On Plus and Pro (and a limited group of Free and Go accounts), credits are a pay-as-you-go balance that kicks in after included usage, billed per token against OpenAI's rate card. They are valid for 12 months, non-refundable except where required by law, and an automatic reload can keep the balance topped up under a monthly spend cap. One sharp edge from the help center: a task that starts with a positive balance can finish after concurrent usage depletes it, leaving the balance negative, and the next purchase offsets the hole first.
- **Buy an instant reset.** Plus and Pro only: refreshes both windows immediately and re-anchors your weekly schedule, as described above. Reaching only the five-hour limit does not trigger the offer banner, but the option can still sit in desktop Usage settings.
- **Use a banked reset.** Some promotions (referrals, launch offers like the September 3 to 4, 2026 Astra resets) grant a saved reset that refreshes both windows when you apply it. Banked resets must be used within 30 days of being granted unless the offer says otherwise.
- **Run on an API key.** API-key usage bills at standard API rates and never touches subscription limits, but it covers local clients only (CLI, SDK, IDE), with no cloud features. As of October 2026 the API lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output (standard, up to 272K context). [Subscription vs API credits](https://aq.dev/guides/subscription-vs-api-credits-for-coding-agents/) covers when each billing shape wins, and [token costs for teams](https://aq.dev/guides/coding-agent-token-costs-for-teams/) covers budgeting it across people.
- **Note what nobody can do:** OpenAI Support does not reset usage limits, for any plan.

## Is it a limit or an outage?

A limit announces itself: the banner or error names the exhausted window and shows a reset time, and /status agrees. An outage looks different: connection failures, 500-series errors, or timeouts that never mention limits. When the error text is ambiguous, check status.openai.com (OpenAI's status page lists Codex by surface: web, API, CLI, IDE extension) and whether teammates are affected; if the status page is clean and your meters read exhausted, it is your account, not the service. [What to do when Claude Code or Codex is down](https://aq.dev/guides/what-to-do-when-claude-code-or-codex-is-down/) is the triage guide for the other case.

## How do Codex limits compare with Claude Code limits?

The shape is similar and the details differ. Both stack a roughly five-hour window under a weekly cap, share the pool across the vendor's surfaces, and sell a pay-as-you-go overflow (OpenAI credits, Anthropic usage credits). The differences, as of October 2026: Codex's five-hour window disappears entirely on Pro tiers and Business Premium seats, while every Claude subscription keeps its session window; OpenAI sells instant resets and grants banked resets, which have no Claude equivalent; and Anthropic's session window opens at your first message and rolls for five hours, while a new Codex window starts with your first message after the previous window ends. [Claude Code usage limits explained](https://aq.dev/guides/claude-code-usage-limits-explained/) is this guide's twin on that side.

## Making the allowance land as finished work

Whatever the plan, the waste modes are the same: a long-lived thread that re-reads a bloated history on every turn, Astra pointed at work Luna could do, and runs that die with a closed laptop after spending their window. The habits that fix it: start big runs right after a reset, keep one task per session in its own git worktree, pick the model per task rather than per habit, and keep the machine running so the budget you spent becomes a finished branch. [Keeping Codex running after closing your laptop](https://aq.dev/guides/keep-codex-running-after-closing-laptop/) and [running Codex on a cloud VM](https://aq.dev/guides/run-codex-on-a-cloud-vm/) cover the persistence half.

## Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. AQ does not change your Codex limits and does not resell model access: each person connects their own ChatGPT account through a per-user CLI login, and AQ never marks up usage on your own subscription. What it changes is how much of the allowance becomes shipped work. Codex runs as the real CLI in a persistent tmux session on your team's VM, so the run you start at the top of a window keeps working after your laptop closes and is resumable from any device. Each workspace gets its own isolated git worktree, which makes fresh-session-per-task the default instead of a discipline, and teammates can open the same workspace and watch or drive the same live session, so when one person's weekly cap runs out mid-feature, a teammate can pick the work up on their own account. The Free plan is a personal sandbox AQ creates for you, nothing to install and no time limit; the AQ Pro plan is $50 per user per month in early access (standard $200, billed monthly), covering VMs you connect from your own cloud or a dedicated always-on AQ-managed VM, with the rate locked for your first 12 months.

## Frequently asked questions

### Does Codex have a five-hour limit?

On some plans. As of October 2026, Plus and Business Standard seats have a five-hour window (OpenAI estimates 15 to 160 local GPT-6.1 Sol messages per window on those plans), while Pro 100, Pro 200, Pro 500, and Business Premium seats currently have no five-hour limit in Work and Codex. Every plan still has an included overall allowance, and a weekly window can apply, so no tier is unlimited.

### How do I check how much Codex usage I have left?

Type /status inside a Codex CLI session: it shows your remaining five-hour and weekly allowance with reset times, and checking is free. The same meters live at chatgpt.com/codex/settings/usage and under Settings then Usage in the ChatGPT apps. If the numbers look wrong, note that OpenAI Support cannot reset limits; it can only investigate miscounted usage.

### What happens if I hit a Codex limit in the middle of a task?

Per OpenAI's documentation, Codex can finish the active turn even after you reach a limit, subject to fair-use limits. After that turn it stops accepting new prompts and the limit banner shows your options: wait for the reset, add credits, apply a banked reset if you have one, buy an instant reset on Plus or Pro, or upgrade. Your conversation and working tree are not lost.

### Do Codex cloud tasks have separate limits from the CLI?

No. Local messages and cloud tasks draw from the same shared Work and Codex allowance, and OpenAI notes cloud tasks may use more of it than local messages. Standard Codex Cloud environments add no separate VM charge as of October 2026, so the allowance is the whole bill. Delegating to the cloud buys persistence (tasks continue while your computer sleeps), not extra budget.

### Can I buy more Codex usage without upgrading my plan?

Yes, two ways on Plus and Pro as of October 2026. Credits are a pay-as-you-go balance billed per token after your included usage runs out; they last 12 months and can auto-reload under a monthly cap. An instant reset refreshes both the five-hour and weekly windows immediately, but it pulls your normal weekly allowance forward rather than adding to it, and your next automatic weekly reset moves to seven days after your first request following the reset. A third path, an API key, bills at API rates outside your subscription entirely but only covers local clients.

### Why did my Codex weekly reset date change?

Almost certainly a reset was applied: purchased instant resets and full banked resets both refresh the weekly window and re-anchor the schedule, with the new weekly period starting at your first Work or Codex request afterward and the next automatic reset seven days after that. This is documented behavior, not a bug. Automatic resets OpenAI grants during incidents or launches (like the September 2026 Astra launch resets) refresh limits the same way.

---

About AQ: AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. Start at [aq.dev](https://aq.dev/).