On this page
If you use Codex through a ChatGPT plan, the number you most want is the one OpenAI doesn't publish. Every plan from Free to Enterprise includes Codex, yet OpenAI gives no weekly allowance for any of them. The closest thing to a hard figure is a set of estimates for how many local messages Plus and Business Standard users can send in a five-hour window, and even those span a tenfold range. What OpenAI does spell out is how quickly your choices eat the allowance. Fast mode counts at 2.5 times the normal rate, Ultrafast at eight times, and the gap between GPT-6 Luna and GPT-6 Astra is enormous. Once you know that, you can make a plan last much longer, or work out when credits or an API key are the better deal.
Everything here comes from OpenAI's Codex pricing and speed pages, its help center and posts by OpenAI staff, last checked on November 8, 2026, Beijing time. These rules change often, so read the figures as a snapshot of this week.
Two windows and one shared pool
According to OpenAI, Codex limits can apply over a five-hour window and a weekly window, depending on your plan. Where both apply, you need room in both to keep going. The five-hour clock isn't fixed to the time of day: a new window starts with your first Work or Codex message after the previous one ends. That is also why you can run out inside five hours while most of your weekly allowance is untouched, which OpenAI describes as expected behavior.
The pool is wider than Codex itself. OpenAI says ChatGPT Work and Codex share the same pricing, credits and usage limits, and that local messages and cloud chats draw from the same allowance. The Codex pricing page adds that on Plus and Pro the limits are also shared with other agentic features once their pricing takes effect, which currently means ChatGPT for Excel. OpenAI's credits article for personal plans goes a step further and says Codex, ChatGPT Work, Excel and PowerPoint can share one agentic allowance on Plus and Pro, with Word drawing on it too. A long spreadsheet session, in other words, can cost you coding time. OpenAI's help center adds Workspace Agents to the shared pool where your plan has them, and says tasks you start through ChatGPT Voice in Work or Codex on desktop, and building or editing a Site, count toward the same Codex usage. The voice connection itself is metered too. According to OpenAI's Codex pricing page, Voice in the desktop app uses the existing Codex usage budget at $0.05 per minute, while the model doing the task is billed separately at its own token rates. On credit-based Business, Edu and Enterprise workspaces, and for Plus and Pro users spending extra credits, desktop voice costs 1.25 credits per minute. Cloud machines, on the other hand, cost nothing extra for now: OpenAI says standard Codex Cloud environments have no separate VM charge at launch, and the model usage counts toward normal limits. ChatGPT's own caps on image generation, file uploads and voice are a separate matter, and OpenAI says those banners don't apply to Codex.
OpenAI is blunt that the allowance isn't a message count. Model choice, context size, reasoning, tool use, retrieval and caching all change what a task costs, so two requests that look alike can land very differently. The only gauge you can trust is your own: Settings and then Usage in ChatGPT or the Codex app, or the /status command inside a Codex CLI session.
What each plan gets
This is how OpenAI describes each plan's Codex allowance today. Where no number appears, OpenAI hasn't published one.
- Free ($0) and Go ($8/month): GPT-6 Luna at Standard speed in the desktop app, subject to rollout. Codex Cloud isn't included. No five-hour or weekly figures.
- Plus ($20/month): GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna under a five-hour window, plus limited GPT-6 Astra usage, which OpenAI estimates at 5 to 45 local messages per five hours. OpenAI says weekly limits may also apply. It's the only personal plan with published five-hour estimates.
- Pro ($100, $200 or $500/month): no five-hour limit for now, though each tier still has an included allowance. OpenAI says Pro $200 includes more than Pro $100 and Pro $500 the most, without giving amounts. Pro $100 and Pro $200 can spend their full allowance on Astra. Only Pro $500 gets Ultrafast. On 2026-11-08 at 13:09 Asia/Shanghai, Tibo (@thsottiaux) posted publicly that, compared to Plus, Pro 100 is 10X, Pro 200 is 20X, and Pro 500 is 50X with access to ultrafast, and that all Pro plans have no 5-hour limits (https://x.com/thsottiaux/status/2119296267545419987). That ladder is from the public post—not wording that showed up on chatgpt.com/pricing or the Pro / usage Help pages as of 2026-11-08 Asia/Shanghai. The 10X / 20X / 50X figures are plan-level multiples versus Plus; they are separate from the Fast 2.5× and Ultrafast 8× speed-mode burn rates covered below. We haven't measured absolute message or token counts on each tier—Settings → Usage and in-product limit notices are the live gauge. More on the Plus-to-Pro price ladder: ChatGPT Plus vs Pro.
- Business ($20 per user per month billed annually, $25 monthly): Standard seats get the same five-hour estimates as Plus, including limited GPT-6 Astra usage of an estimated 5 to 45 local messages and GPT-6.1 Terra at an estimated 35 to 285. Premium seats, at $100 a month billed annually or $125 monthly, get five times Standard's included usage and no five-hour limit, according to OpenAI.
- Enterprise and Edu: workspaces on credit-based flexible pricing have no fixed rate limits; usage scales with the workspace's credit pool. Without flexible pricing, OpenAI says the per-seat limits match Plus for most features.
For Plus and Business Standard, OpenAI estimates these local-message ranges per five-hour window:
- GPT-6 Astra: 5 to 45
- GPT-6.1 Sol: 15 to 160
- GPT-6.1 Terra: 35 to 285
- GPT-6 Sol: 15 to 150
- GPT-6 Luna: 350 to 3,000
OpenAI calls these estimates rather than guarantees, and says cloud tasks can use more than local messages. The ranges are wide enough that the comparison between rows matters more than any single figure: at the bottom of each range, Luna gives you 350 messages where Terra gives you 35, Sol gives you 15 and Astra gives you five. GPT-6.1 Terra is available in Work and Codex on Plus, Pro, Business, Enterprise and Edu plans, not in regular Chat; for reference only on how many messages a real session uses. Related reading: GPT-6.1 Terra. One inconsistency is worth knowing about. OpenAI's Business help article groups GPT-6.1 Sol with GPT-6 Sol at 15 to 150, while the Codex pricing page and OpenAI's Astra usage guide put GPT-6.1 Sol at 15 to 160. We've gone with the pricing page. The same Business help article also gives Standard seats five-hour estimates for older models, 10 to 100 local messages for GPT-5.6 Sol, 25 to 200 for GPT-5.6 Terra and 250 to 2,000 for GPT-5.6 Luna, figures the Codex pricing page doesn't list.
There's a second inconsistency at the free end. The Codex pricing page says Free and Go get GPT-6 Luna at Standard speed in the desktop app, subject to rollout. OpenAI's ChatGPT pricing page, however, lists GPT-5.6 Terra with "Limited access in Work and Codex on desktop" for both plans, and a help center article on ChatGPT's models still names Terra as the Codex model for Free and Go. We've followed the Codex docs, but the model picker on your own account is the final word.
One model has left the lineup since this guide first ran. GPT-5.5 retired from Codex for users signed in with ChatGPT at 03:00 on October 15, Beijing time. OpenAI says the API and Codex sessions signed in with an API key aren't affected, and recommends GPT-6.1 Sol where a plan and client support it. Our story on the GPT-5.5 retirement covers what to switch to.
If you pay for Pro $200, one date is worth circling. OpenAI says new Pro $200 subscriptions come with a lower allowance than before, "to reflect our increasingly efficient models." Anyone with an active Pro $200 subscription at any point from September 22 through 10 a.m. Pacific on September 29 keeps the previous allowance through October 29, then moves to the lower one at the same $200 price. OpenAI's help article adds that the previous allowance applies only while the Pro $200 subscription is active. OpenAI hasn't said how much lower it is. Keeping the old allowance also doesn't upgrade the plan or add Ultrafast, according to the same article.
Why Astra and Fast mode drain you
The credit rate card shows the cost gap between models most clearly. At Standard speed, per million tokens, OpenAI charges:
- GPT-6 Astra: 250 credits input, 25 cached input, 1,250 output
- GPT-6.1 Sol: 50 credits input, 2.5 cached input, 250 output
- GPT-6.1 Terra: 25 credits input, 1.25 cached input, 125 output
- GPT-6 Sol: 50 credits input, 5 cached input, 250 output
- GPT-6 Luna: 2.5 credits input, 0.25 cached input, 12.5 output
So token for token, Astra costs five times as much as Sol, Sol costs twice as much as Terra, and Terra costs ten times as much as Luna. OpenAI warns that credit prices alone don't determine your included usage, but the message estimates above point the same way.
Speed multiplies all of this. OpenAI's Speed page says Fast mode works on GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna where available, and uses included subscription limits at 2.5 times the Standard rate, while purchased credits and Enterprise pay-as-you-go usage are billed at twice the Standard rate. Ultrafast supports only GPT-6 Astra and GPT-6.1 Sol. It uses included limits at eight times the Standard rate and bills credits at six times. OpenAI says GPT-6 Astra Ultrafast generates tokens up to 8x faster than Astra in Standard mode in Codex. Ultrafast is limited to Pro $500 and eligible Enterprise and Edu plans, and OpenAI says buying credits on any other self-serve plan doesn't unlock it. On Pro $500, it spends your included usage first and then your credits. Our news story "OpenAI adds Ultrafast mode to GPT-6.1 Sol: up to 8x faster, at six times the price" covers the API side.
Notice that both speed modes hit your plan harder than your credit balance. Fast costs 2.5x against included usage but 2x in credits; Ultrafast costs 8x against one and 6x against the other. Put Fast mode on Astra and you are combining the most expensive model with a 2.5x multiplier, which on Plus is the quickest route to an early end to your five-hour window.
A few other things move the needle. OpenAI says image generation uses included limits three to five times faster on average than similar turns without images, and isn't available on the Free plan at all. Higher reasoning effort uses more allowance and "does not always produce a better result." OpenAI also says Astra at Low effort can outperform Sol at High, so stepping down effort on a stronger model is worth trying before stepping up on a weaker one. Every MCP server you leave enabled and every extra line in a large AGENTS.md adds context to each message, and OpenAI's own advice is to trim both. One recent change works in your favor: auto-review safety checks are free when you're signed in with ChatGPT and don't count toward plan limits. Tibo (@thsottiaux) wrote on October 7, Beijing time, that auto-review can account for 2 to 10 percent of a plan when used.
When you hit the wall
If you reach a limit in the middle of a turn, OpenAI says Codex can finish that turn, subject to fair-use limits. What happens next depends on your plan.
On October 21, Beijing time, Tibo (@thsottiaux) wrote that over the weekend some users reported usage draining faster than anticipated. About 40,000 accounts had an inconsistency where the usage limits that were actually applied did not match the selected plan. Tibo said the issue has been corrected, that those users are being contacted individually, and that affected accounts will receive credits.
Update (October 21, Beijing time): Tibo, who works on Codex at OpenAI, said in a post on X that for logged-in users across Codex and Work, OpenAI keeps the cache forever. According to the post, there is no penalty for coming back to an old conversation or resuming after a few hours, and those requests are now guaranteed to hit the cache. This comes from Tibo's post, not from OpenAI's documentation or changelog. The same post mentions the new 12-hour prompt caching option in the API, which is a separate API setting.
On Plus and Pro, you can buy credits and keep going without upgrading. Your included usage is always spent first, and only then does Codex draw on credits. You can buy them, or turn on automatic reload, from Settings and then Usage in ChatGPT, or Usage & Billing in the Codex app. Automatic reload tops your balance back up to a target you choose, and OpenAI says you can cap it with a maximum monthly spend, which applies to reloads but not to one-time purchases. The same balance works across Codex, ChatGPT Work, Word, Excel and PowerPoint. Credits are valid for 12 months and are non-refundable except where law, contract or an approved exception says otherwise, and they aren't API credits. A task that started with a positive balance can push it negative if other usage runs at the same time. What a credit costs in dollars isn't on the pages we checked; OpenAI says purchase prices and discounts depend on your plan or agreement. Free and Go are mostly shut out for now: OpenAI says only a limited group of those users can buy credits, and doing so doesn't raise the plan's included limit.
Plus and Pro accounts may also see an option to buy an instant reset, depending on account and billing country. It isn't offered on Free, Go, Business, Enterprise or Edu. There are two ways in: the Usage settings in the ChatGPT desktop app, where the option appears once some of the five-hour or weekly allowance has been used, or an in-app banner after hitting the weekly limit. OpenAI notes that reaching only the five-hour limit doesn't bring up that banner, though the settings option may still be there. It restores both the five-hour and weekly allowance the moment checkout succeeds, even if you had usage left, and it can't be saved for later. OpenAI is clear that it pulls your normal weekly allowance forward instead of adding extra usage: your next automatic weekly reset moves to seven days after your first Work or Codex request following the purchase. OpenAI doesn't list the price outside checkout and says it generally doesn't refund these purchases. If you still have allowance left and your regular reset is close, waiting is often the cheaper move.
Then there are banked resets, the free ones OpenAI hands out as promotions or referral rewards. A full banked reset refreshes both windows and also moves your weekly reset date, and it's only used up if it actually refreshes something. They arrive often enough lately that it pays to check Settings and Usage before buying anything. At 03:19 on October 8, Beijing time, Tibo posted: "Loading a banked reset in everyone's paid accounts." OpenAI says future resets aren't guaranteed and that Support won't reset limits for you.
Between October 13 and 18, Beijing time, OpenAI's Tibo Sottiaux announced three Codex usage resets on X. They were one-off moves, not a change to the limit rules in this guide: OpenAI says a reset restores your allowance without permanently increasing your plan's limits, and that future resets are not guaranteed. Our news story on Tibo's three resets this week has the dates, times and what he said about them. On Oct 22 (14:55 Beijing time), Tibo announced the Day 17 reset; per his reply it "Lands tomorrow before 10am PST" (window ending before about 02:00 Beijing time on Oct 23). Accounts report the reset credited within that window. Tibo has not posted a separate Done confirmation as of this update. Tibo's announcement
On Oct 23 (12:46 Beijing time), Tibo announced the Day 18 reset; per his reply it is "Landing tomorrow before 10am PST together with a small ship" (window ending before about 02:00 Beijing time on Oct 25). Accounts report the reset credited as of about 01:50 Beijing time on Oct 24, within that window. Tibo has not posted a separate Done confirmation as of this update.
On Oct 25 (11:39 Beijing time), Tibo announced the Day 20 reset with "Day 20/ Rest and reset"; per his reply it is "By midnight PST" (window ending before about 16:00 Beijing time on Oct 25). At about 14:06 Beijing time he confirmed it had landed ("It is all propagated"). Confirmed live by about 14:06 Beijing time.
Accounts report the Day 21 reset credited as of about 16:01 Beijing time on Oct 26. Tibo has not posted a separate Done or propagated confirmation as of this update.
On Oct 28 (10:42 Beijing time), Tibo announced the Day 23 reset; per his reply it is "Lands by midnight PST" (window ending before about 16:00 Beijing time on Oct 28). At about 14:29 Beijing time he confirmed it had landed ("And it's done. See you tomorrow for some good news."; https://x.com/thsottiaux/status/2115330127886463244). Accounts report the Day 23 reset credited as of about 14:53 Beijing time on Oct 28.
On Oct 30 (00:56 Beijing time), Tibo announced Day 25: doubling all usage limits for the Pro 100, Pro 200 and Pro 500 tiers and for Business Premium, effective that day (https://x.com/thsottiaux/status/2115850175868727621). The same post said Pro 200 limits would not be reduced, so Work and Codex usage stay the same; extended limits for dots continue for now, and dot usage remains separate from default usage. On 2026-11-08 at 08:54 Asia/Shanghai, Tibo replied that “We doubled the usage on Pro plans and Business 100”—past tense pointing at the Day 25 doubling already on this page, not another round of doubling (https://x.com/thsottiaux/status/2119232059537514598). That reply said “Business 100”; Day 25 wording here stays with Business Premium for the about-$100 annual Premium seat.
Business works differently. Members use their included seat limits first, then draw on the workspace's shared credit pool if an owner has bought credits. With no credits in the pool, the feature is blocked until an owner adds more. Usage-based Codex-only seats need workspace credits from the first message. Enterprise and Edu workspaces on flexible pricing share a contract-level pool, and once it's gone, advanced features pause unless owners allow overages or buy more. Team Tasks in ChatGPT also draw on the workspace, and OpenAI's documentation says: "Team Tasks use workspace credits. Team spending limits are separate from user limits."
OpenAI Help also previews that starting in December, Plus, Pro, and Business usage allowances will be shared across ChatGPT and Codex and measured in tokens (model and task complexity affect burn); Instant chat will not draw from that shared usage, and combined remaining usage and reset times are planned under Settings → Usage (https://help.openai.com/en/articles/20001625-understanding-chatgpt-and-codex-usage-limits). The change isn't live yet—no start day, timezone, or per-plan token totals—so the plan figures above stay as they are.
The API key route
There's one exit from plan limits altogether. You can sign in to the ChatGPT desktop app, the Codex CLI or the IDE extension with an OpenAI API key, and OpenAI then bills you through your Platform account at standard API rates. Plan windows and ChatGPT credit multipliers no longer apply. The trade-offs are real, though. You lose cloud features such as Codex Cloud, GitHub code review and the Slack integration, and the models you can use are whatever your API key has access to. OpenAI says any user can run extra local chats this way and recommends API keys for automation such as CI jobs.
At short context, the API list prices per million tokens are $2 input, $0.10 cached input and $10 output for gpt-6.1-sol; $0.10, $0.01 and $0.50 for gpt-6-luna; and $10, $1 and $50 for gpt-6-astra. On OpenAI's API pricing page, Fast is exactly twice Standard, GPT-6.1 Sol Ultrafast is $12 input and $60 output, and GPT-6 Astra Ultrafast is $60 input and $300 output. OpenAI also cautions against using API token prices to estimate how many tasks your subscription includes, since the two are metered separately.
Which limit you keep hitting decides the plan
Watch which meter runs out first. If it's the five-hour window, that is the one thing Pro and Business Premium seats remove, though both still carry an included allowance. If it's the weekly limit, start with model choice: moving routine work to GPT-6 Luna and saving Astra for the problems that need it does more than any setting, and turning off Fast mode on long jobs is the next easiest win. If what's eating your allowance is a scheduled script or CI job, an API key takes it out of the shared pool. Plus users who only go over occasionally have credits and instant resets to bridge the gap. The jump from Plus at $20 to Pro at $100 makes the most sense if you hit the wall every week, with the caveat that OpenAI hasn't said how big Pro's allowance is, beyond the fact that each higher tier gets more.
