Codex Usage Limits: How to Check Them and Make Them Last

Codex on a ChatGPT plan does not give you a fixed number of messages. You get an allowance that resets every five hours, with weekly limits that may also apply, and every Codex surface draws on it: the CLI, the IDE extension, the ChatGPT desktop app and cloud tasks, plus ChatGPT Work. How far it goes depends on the plan, the model you choose and how much each task makes the model read and write. To see where you stand, type /status in a Codex CLI session or open the usage dashboard in Codex settings. When it runs out, Plus and Pro users can buy credits, workspaces on flexible pricing can buy workspace credits, and anyone can keep working locally with an API key at API rates.

How the limits are counted

OpenAI’s Codex pricing page↗ (opens in a new tab) explains what moves the meter. The points that matter in practice:

• One allowance for everything. Local messages and cloud chats share it, and ChatGPT Work inside ChatGPT uses the same limits and credits as Codex.
• Windows, not a monthly total. Estimates are given per five-hour period, and weekly limits may also apply.
• Size, not count. A small script uses a fraction of what a long session holding a lot of context uses. Model, reasoning, tool use, retrieval and caching all change the cost, so prompt length alone is a poor guide.
• Cloud costs more. Cloud chats on ChatGPT plans use GPT-5.6 Sol and may use more of the allowance than local messages.
• Extras cost more. Fast mode and other speed settings use limits faster, and image generation uses them three to five times faster on average.
• Code review on GitHub is counted as Code Review usage. A review you run locally counts against the general limit.

The figures OpenAI publishes

OpenAI gives ranges, not promises: estimates of local messages per five-hour window, which it says are not fixed limits. As listed on 28 September 2026:

• Plus: 15 to 150 local messages with GPT-6 Sol, 350 to 3,000 with GPT-6 Luna, and 5 to 45 with GPT-6 Astra.
• Pro: two tiers with 5x or 20x the Plus usage. Pro 5x, for example, lists 70 to 700 messages with GPT-6 Sol; Pro 20x lists 300 to 3,000.
• Standard Business: the same estimates as Plus.
• Enterprise and Edu with flexible pricing: no fixed rate limits; usage draws on credits. Without flexible pricing, the same per-seat limits as Plus for most features.
• Free and Go: limited access for quick tasks, with GPT-6 Luna in the desktop app as it rolls out. No message estimates are published.

The ranges are wide because tasks are. One more date to note: OpenAI says GPT-5.5 retires from ChatGPT and Codex on all plans on 14 October 2026, so a saved configuration that names gpt-5.5 needs changing before then.

Where to see what you have left

• /status in the Codex CLI: the active model, approval policy and token usage for the session. OpenAI’s pricing FAQ names it as the way to see remaining limits during a session.
• /usage in the CLI: account token activity, with /usage daily, /usage weekly and /usage cumulative for each view. The same menu redeems an earned rate-limit reset if you have one.
• /statusline: puts items such as the model, context and limits in the CLI’s footer, so you see them without asking.
• The usage dashboard, under Codex settings in ChatGPT: current limits and reset times. OpenAI suggests checking it every week or two to understand your pace.
• For workspaces: admins have usage insights and spend controls across ChatGPT Work and Codex.

/status              # model, policy and token usage for this session
/usage weekly        # account token activity for the week
/statusline          # add limits and context to the footer
/compact             # summarise a long chat to free context
/new                 # start a fresh chat for the next task

The slash commands reference↗ (opens in a new tab) has the full list. If /usage asks you to sign in, the session is not signed in with ChatGPT, for example because it runs on an API key, and plan limits do not apply to it.

When you hit the limit

Codex does not cut you off mid-thought: if you reach the limit during a turn, the agent can finish that turn, subject to fair use. After that you have three options. Plus and Pro users can buy credits and carry on without changing plan. Business, Edu and Enterprise workspaces on flexible pricing can buy workspace credits. And anyone can run extra local chats with an API key, charged at standard API rates. An API key covers the CLI, the SDK and the IDE extension, but not cloud features such as cloud tasks, GitHub code review or Slack. OpenAI’s help-centre articles on buying credits could not be checked for this page, so read the prices and discounts on your own plan’s page.

Habits that make the allowance last

OpenAI’s own list comes first: keep prompts precise, give only the relevant files, say what output you need, trim AGENTS.md, and switch off MCP servers you are not using, because every server adds context to every message. These go further:

• Match the model to the job. OpenAI’s models page↗ (opens in a new tab) recommends GPT-6 Sol for complex coding and agent work and GPT-6 Luna for focused, repeatable tasks. On the published estimates, Luna goes many times further per window.
• Leave reasoning effort at the default. Higher effort takes longer and uses more tokens; raise it for the task that needs it, then put it back.
• Split AGENTS.md by folder. Codex reads the files from the repository root down to the folder you start in, so rules kept in a service’s own folder stay out of sessions started elsewhere.
• Run small jobs locally. A cloud task on a ChatGPT plan uses GPT-5.6 Sol and may cost more than the same job in the CLI.
• Ask for several cloud attempts only when you will compare them. Each attempt is its own run.
• Start fresh between tasks. /new drops the previous task’s history; /compact shrinks a long session you need to keep.
• Leave Fast mode off unless speed matters more than allowance.

The broader habits for a Codex session, such as scoped tasks, approval modes and review, are in Codex CLI best practices. The Claude equivalent of this page is how to reduce Claude Code token usage.

Scoped tasks cost less

Most wasted usage is the agent working out what you meant: reading half the repository, trying one approach, then another. A task that already says what, why and where, with a plan and a definition of done, saves that exploration. That is what a fenbs task holds: a note with the problem and the file, a plan with the steps, and a test status with notes when it is finished. Codex reads it over MCP, works it, and records the result in History under its name. Follow OpenAI’s advice on MCP here too: in config.toml, list only the fenbs tools a session needs in enabled_tools, so the board adds as little context as possible.