Codex Usage Limits Explained: The 5-Hour Window, the Weekly Cap, and What Actually Resets (2026)

TL;DR: Codex Usage Limits Are a Window, Not a Count

The question "what are the Codex usage limits" has no single-number answer, and that is by design. OpenAI's pricing page, read today, describes an allowance that refills on a rolling five-hour window, varies by which model you run, may also be capped weekly, and is spent by tokens rather than by messages. Plus is the base allowance and a standard Business seat matches it; a Premium Business seat gets five times Standard with no 5-hour limit, and Pro ($100, $200 or $500) currently has no five-hour limit either. Nothing here is a fixed cap, which is why two people on the same plan report wildly different experiences. One runs Luna for quick edits and never sees the limit; the other keeps a 40-file repo in context on Sol and hits it before lunch.

I run Codex, Claude Code, and Cursor side by side on real client work, and the honest version of this post is that the limits stopped being the interesting question once I understood the meter. Here is the meter.

The Codex limits by plan, from OpenAI's own page (September 2026)

These are OpenAI's published estimates of local messages per five-hour period, re-read on September 29, 2026, and the page says out loud that they are estimates, not fixed message limits. Cloud tasks may consume more of the allowance than a local message does. A standard Business seat gets exactly the Plus column. Pro no longer has a column: the page says "Pro plans currently have no five-hour limit."

Last checked October 12, 2026: every range in this table still matches OpenAI's page, and Pro still has no five-hour limit. Two things were added since the September read, both further down this post: GPT-6.1 Sol now has an Ultrafast mode at the same 8x as Astra, and ChatGPT for Excel on Plus and Pro now draws from this same allowance.

The GPT-5.6 rows are gone from that table as of the September 29 read (GPT-5.6 still has credit rates further down the page), and GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on every plan on October 14, 2026. OpenAI's Business help article still lists GPT-5.6 Sol at 10–100 for a Standard seat, and groups GPT-6.1 Sol with GPT-6 Sol at 15–150, so the two pages disagree by ten messages at the top of the 6.1 range.

Three lines of fine print matter more than the table. A standard Business seat ($20 per user per month billed annually, $25 monthly, two-user minimum) gets the Plus estimates; OpenAI's Business article says "Premium seats have 5x the included usage of Standard seats, with the benefit of no 5-hour limit." "Weekly limits may also apply" is the page's exact phrasing, with no weekly number published. And "ChatGPT Work and Codex share usage", meaning agentic work you run inside ChatGPT draws down the same allowance as your CLI sessions. As of October 12, 2026 the page adds ChatGPT for Excel on Plus and Pro to that shared pool, so a long spreadsheet session spends the same budget as your coding. The Go tier ($8/month) exists for lightweight coding and is not in the estimate table; the API key route has no plan windows at all and bills at usage rates instead.

One Astra message, in other models' messages

OpenAI never says how the models compare, but its own ranges do. Divide each model's range by GPT-6 Astra's on the same plan and you get an exchange rate for the window. The ratios hold for a standard Business seat too, because its column is identical to Plus.

A worked morning, to make that concrete. Say you send ten Codex turns on Plus. On Astra, ten turns are 22% of a window if your tasks are small enough to sit at the top of OpenAI's range (45 per window), and two full windows if they're big enough to sit at the bottom (5 per window), which is ten hours of waiting. The same ten turns on GPT-6 Sol are 7% to 67% of one window. The assumption doing the work: a task that sits at the top of one model's range sits at the top of the others' too. OpenAI doesn't state that, but it is the only way to read ranges published side by side.

The practical rule it gives you: each Astra turn costs about three GPT-6 Sol turns, so keep it for the turn that needs it and let GPT-6 Sol or Luna do the rest.

Codex on Free and Go, and the "monthly limit" question

Free ($0) and Go ($8 a month) are both listed as "GPT-6 Luna at Standard speed in the desktop app, subject to rollout." Neither gets a column in the estimate table, so OpenAI publishes no per-window number for either; the web, CLI, IDE and iOS surfaces and the cloud integrations are listed on the Plus card, not on these two. Image generation isn't available on Free at all.

There is no monthly Codex limit on a ChatGPT plan. The two clocks are the five-hour window and the weekly cap. The only monthly-shaped meters are on the business side: Enterprise and Edu workspaces with flexible pricing have "no fixed rate limits" and spend credits instead, and without flexible pricing they get "the same per-seat usage limits as Plus."

How to check your Codex usage (and what uses it up)

There are two places to look, and OpenAI names both on its Codex pricing page (read October 10, 2026). The usage dashboard at chatgpt.com/codex/settings/usage shows your current limits and when each one resets. The /status command shows what's left from inside an active Codex CLI session, which is the one to run before you start a long task, not after it stalls.

What you'll see is a share of an allowance, not a message count, because OpenAI's numbers are estimates: 15 to 160 GPT-6.1 Sol local messages per five hours on Plus and on a Standard Business seat, and no five-hour limit on Pro plans. Where you land in that range depends on what you ask for. Three things move it most:

• Speed mode. Fast uses your included usage at 2.5x the Standard rate. Ultrafast on GPT-6 Astra or GPT-6.1 Sol uses it at 8x.
• Image generation. 3 to 5x faster than a similar turn without images, depending on quality and size.
• Cloud tasks and big context. Cloud tasks may use more than local messages, and long sessions that hold a lot of the codebase in context cost more per message than small scripts.

OpenAI's own advice is to check the dashboard every week or two to learn your pace. Mine is the same with a deadline: three normal working days, then decide whether you need more room or a smaller model.

The 5-hour window: what resets and when

The window is rolling, not calendar-based. Usage you spent at 9am frees up at 2pm regardless of what you did in between, which is why a heavy morning and a heavy afternoon feel like two separate budgets and a heavy single stretch feels like a wall. The ChatGPT usage dashboard shows current limits and reset times; inside a Codex CLI session, /status shows the same thing without leaving the terminal. If you have never typed that command, type it before you decide you need a bigger plan.

The Codex weekly limit

OpenAI states that a weekly cap may apply on top of the five-hour allowance and does not publish it, which is the single most-searched frustration about this product and the reason "codex weekly limit" is its own query. Read the two meters together and the order of events is predictable: on an ordinary week the five-hour window is the one you meet, and the weekly cap only shows itself after several days of long agentic runs in a row, at which point the five-hour reset stops helping. If you are hitting it, you are a heavy user by any definition, and the plan question below is a real one for you rather than a hypothetical.

Codex rate limits vs usage limits

The two phrases get used for the same thing, and the page itself mixes them. The rolling window and the weekly cap are what a ChatGPT-plan user means by either. The other meaning, rate limits in the API sense (requests and tokens per minute by API tier), only applies if you authenticate Codex with an API key instead of a ChatGPT login. That route drops the plan windows entirely, bills at API pricing, and gives up the cloud features (GitHub code review, Slack) that live on the ChatGPT side. It is the right choice for CI and shared environments and the wrong one for a person who wants a predictable monthly bill.

What burns the allowance faster than you think

• Context. A session holding a large codebase costs more per turn than a fresh one; the page lists context, reasoning, tool use, retrieval, and caching as the variables, and says prompt length alone is not a reliable estimate.
• Cloud chats. OpenAI says cloud tasks may use more of the allowance than local messages.
• Fast mode. Fast uses included subscription usage at 2.5x the Standard rate (2x on purchased credits), and Ultrafast at 8x (6x on credits) on both GPT-6 Astra and, as of the October 12 read, GPT-6.1 Sol. OpenAI notes those are billing multipliers, not speed increases.
• Images. Image generation draws on the same limits roughly three to five times faster on average.
• Voice. Voice in Desktop now spends your existing Codex budget "at $0.05 per minute," and the model doing the task is billed on top at its normal rate. On credit-billed workspaces it's 1.25 credits a minute. (Earlier this month the page gave Voice its own allowance and listed a separate Codex-Spark limit; neither appears on the September 28 read.)

What happens when you hit the limit mid-task

The page is explicit that work in progress gets to finish: if you reach the limit during an active turn, the agent continues that turn subject to fair use. So a long migration does not die halfway; it stops accepting the next instruction. From there you have four moves, cheapest first: wait for the window, switch to a smaller model (Terra or Luna instead of Sol) so what remains lasts longer, buy additional credits (Plus and Pro can do this without changing plan), or run extra local chats against an API key at usage rates. Upgrading is the fourth option, and it is the one people reach for first.

What is Luna Reserve (the "GPT Reserve" limit) in Codex?

If your usage screen shows a second meter called Luna Reserve, or a model labeled gpt-reserve, that is OpenAI's fallback allowance, not a new plan. OpenAI's help article on it (read October 16, 2026) says it gives selected accounts extra usage on GPT-5.6 Luna once regular usage runs out. Four details matter more than the name:

• Personal Plus and Pro accounts only. Business and Enterprise workspaces don't get it. So the same developer can "last longer" on a home Plus account than on the company's Business seat, and the plan isn't the reason.
• It doesn't refill your regular allowance. The five-hour window and the weekly limit above keep their own clocks. The reserve is a separate allowance with its own meter, and OpenAI doesn't publish its size.
• It doesn't unlock the stronger models. You get Luna, the small model. By the exchange rates above, a Luna turn is the cheapest turn in the window. That's also why it can't stand in for Sol or Astra on a hard refactor.
• Eligibility can change. "Selected accounts" means you might not have it at all, and having it this week is no promise for next week.

What I'd do with it: let the reserve handle the cleanup (renames, tests, small edits) after the window runs out, and save the big task for the reset. If your team runs on Business seats and keeps hitting the wall, the reserve won't help you. The options are the four moves in the section above.

Should you upgrade to Pro for the limits?

Decide it from the meter, not from a bad afternoon. OpenAI's page itself suggests checking the usage dashboard "every week or two" to learn your pace. Run /status at the end of three normal working days and note how often you were actually blocked versus annoyed. If Plus blocks you most days, Pro at $100 is the honest tier, and it is worth knowing that OpenAI now sells that middle rung; the "$20 then $200 cliff" I described in my Codex pricing breakdown has a $100 step in it as of September 2026, and Pro drops the five-hour window entirely. If you are blocked once a week, buy credits and keep the $20 plan. If you are blocked because every session starts by re-explaining the whole codebase, fix the AGENTS.md and the task scoping first; how to use Codex covers both, and it is free. The same three-step order (threads, scoping, then money) is what I wrote for the Claude side in Claude usage limits. If you're weighing a switch rather than an upgrade, Claude's, Gemini's and Copilot's ceilings sit next to OpenAI's in the AI usage limits by plan table.

Codex rate limits: what resets and when

"Codex rate limits" is the phrase people type. The thing they are hitting is the plan allowance above, so here is the reset behavior in one place, from OpenAI's Codex pricing page and its Codex help article, both read on September 15, 2026. There are two clocks. The five-hour window refills on its own schedule, and the usage dashboard shows the exact reset time. The weekly limit resets on a fixed date; OpenAI's help center confirms both exist by describing how a banked reset "refreshes your 5-hour and weekly Codex usage windows and changes your weekly reset date." OpenAI publishes neither the weekly number nor a reset day.

What hitting one looks like: OpenAI publishes no exact error string. In the app you get a limit banner that names which allowance is exhausted and lists your options: add credits, apply an available reset, upgrade, or wait for the displayed reset time. In the CLI, /status shows token usage and /usage shows daily, weekly, or cumulative activity and lets you redeem an earned reset. Support will not reset a limit for you; the help center says so directly.

Two things that are not this limit. API rate limits (requests and tokens per minute by tier) belong to an API key and only apply if you signed in with one. And ChatGPT's own caps on images, uploads, and voice have separate banners and, in OpenAI's words, "do not apply to Codex."

Codex limits when you're paying for the team

If you own the company rather than write the code, here is the version I give on calls. A developer hitting Codex limits every day is either doing genuinely heavy work, in which case $100 a month is the cheapest engineer-hour you will ever buy, or is running the same kind of task by hand every morning, in which case the limit is a symptom and the fix is a workflow, not a plan tier. The AI Readiness Assessment ($2,500, credited in full against a build within 90 days) is how I tell those two apart in two weeks. And whatever tier you land on: before real users touch what Codex shipped, run it against the free 20-point vibe coding security checklist. Unreviewed agent output under a generous allowance is exactly how apps end up in my rescue queue.

Related: Codex pricing · is Codex free? · how to use Codex · Codex review · Cursor vs Claude Code vs Codex · Claude usage limits.

Pro 200 is back, with a smaller allowance

Updated September 29, 2026: the Pro $200 pause I noted on September 20 is over. OpenAI's Pro help article now says "Pro 200 is also available for new subscriptions again," and adds Pro 500 at $500 a month, the only Pro tier with Astra Ultrafast. The catch is in the next sentence: new subscriptions that aren't grandfathered "include a lower usage allowance." If your Pro 200 was active at the cutoff (or in the seven days before it), you keep the old allowance through October 29, 2026, then move to the lower one at the same $200. OpenAI doesn't put a number on either allowance, only that Pro 200 includes more than Pro 100 and Pro 500 the most.

Sources

• OpenAI, Codex pricing (now served at learn.chatgpt.com/docs/pricing; plans, per-model estimates, Pro's missing five-hour limit, speed multipliers, weekly-limit and fair-use wording, Voice, credits): read September 8, 15, 28 and 29 and October 12, 2026
• OpenAI Help, About ChatGPT Pro tiers (Pro 100, 200 and 500; Pro 200 reopened; grandfathered allowance): read September 20 and 29, 2026
• OpenAI Help, ChatGPT Business models and limits (Standard and Premium seat estimates): read September 29, 2026
• OpenAI Help Center, Codex usage and resets article (banked resets, no Support resets, /status and /usage): read September 15, 2026
• OpenAI Help, Luna Reserve in Codex and ChatGPT Work (fallback on GPT-5.6 Luna, personal Plus and Pro only, separate allowance, no regular-usage refill): read October 16, 2026