Codex Rate Limit Reset: When Limits Reset, Weekly Caps, and Credits

Codex usage limits reset automatically on a rolling seven-day clock anchored to your first Codex task of the week, so there is no fixed reset weekday and two developers on the same ChatGPT plan can hit their caps on different days 1. For ChatGPT Plus subscribers the weekly cap is no longer the only meter: a rolling five-hour window runs underneath it again after a six-week suspension, so a Plus account answers to two limits at once 2 3. And since early September the reset itself is something you can buy: eligible Plus and Pro accounts can purchase an instant reset that restores both windows at once, while OpenAI spent the GPT-6 Astra rollout crediting paid users with a free banked reset for every day they waited without the new model 1 4.

Your Codex week starts when your first task does

The weekly window closes seven days after your first request in Codex or ChatGPT Work, and a redeemed reset restarts that same anchor: once you redeem it, your next automatic reset lands about seven days after the redemption 1. A purchased reset behaves the same way: the new weekly period starts with your first request after the reset is applied, and the next automatic reset is scheduled seven days after that request, not seven days after payment 1. The reset time is not a clock rule you can work out at home. It arrives from the server as a resetsAt timestamp, and the honest way to read yours is the /status command inside Codex CLI; the current CLI also documents /usage, which shows token activity and any earned-reset actions available to your account 5. Users have asked OpenAI for a fixed reset day to make the cycle predictable, and separate reports describe the refresh date moving between sessions, which fits a rolling anchor rather than a calendar rule 6 7 8. One GitHub thread documents the sharpest version of the pain: a surprise reset moved a user from 60 percent remaining on the old period to a fresh 100 percent with a new end date five days later, wiping out a surplus the user had budgeted their week around 8.

OpenAI's help center now closes the door users kept knocking on: Support does not reset ChatGPT or Codex usage limits, one-time resets are occasional and eligibility-dependent, future resets are not guaranteed, and the company will not compensate you for a reset you were not eligible for or a banked reset you let expire 9. Community reports said the same before the documentation caught up, and the only fix on record is system-wide: when a bug made redeemed resets vanish without restoring usage, OpenAI applied a system-wide fix with automatic replacement resets rather than adjustments to individual meters 10.

The five-hour window is back on Plus, so two meters can stop your session

OpenAI suspended the rolling five-hour window on 12 July 2026 and called the suspension temporary, for Plus, Business, and Pro accounts 11. The same day, Codex product lead Tibo Sottiaux framed it as a load decision: the previous 48 hours of Codex and ChatGPT Work had been "intense," so OpenAI removed the five-hour restriction for Plus, Business, and Pro, issued a one-time usage reset to everyone, and rolled out changes to make GPT-5.6 Sol consume less of your allowance 12. For Plus it proved temporary in both directions: OpenAI restored the five-hour cap for ChatGPT Plus on Codex and ChatGPT Work on 25 August 2026, ending roughly six weeks of unrestricted access 2 3. The restoration reporting names Plus accounts, while the tracker documentation still lists the window as suspended for Business and Pro; after the 29 September Pro relaunch, OpenAI's Work and Codex pricing guide confirms the Pro side in its own words, reporting that Pro plans currently have no five-hour limit in Codex, only a weekly allowance 11 13 14. The one wrinkle remains: OpenAI's reset-purchase article, which serves Plus and Pro alike, describes a purchase restoring both the five-hour and weekly usage, which reads as the shorter meter mattering on Pro again 1. For Plus the meter is now officially quantified: OpenAI's pricing tables estimate 15 to 150 local messages per five hours on GPT-6 Sol and 350 to 3,000 on GPT-6 Luna, against 10 to 100 and 250 to 2,000 on the GPT-5.6 pair, and the same tables print a Business baseline at the Plus level, which sits oddly with the suspension note above 15. The two windows are tracked independently, which produces Codex's most confusing failure mode: a session can show healthy five-hour capacity while the weekly window is already empty 5. A scheduled weekly reset refills the weekly window on its own clock, but the resets you control reach further: OpenAI's documentation says using a banked reset refreshes both your five-hour and weekly Codex usage windows and changes your weekly reset date, and a purchased instant reset restores both immediately 9 1.

What each plan's cap actually is

OpenAI publishes multipliers, ranges, and averages rather than a fixed weekly credit count for individuals. One metering fact the help center now states in its own words: "Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents use a shared allowance and credit pool when those features are available on your plan," so research tasks in Work spend the same budget as coding runs in Codex 9. What the documentation establishes:

• Free and Go: Codex is included across ChatGPT plans including Free and Go, but those users are prompted to upgrade to Plus rather than buy credits when they hit limits 9 16. Their ladder shifted in late September: GPT-6 Luna is rolling out to Free and Go accounts in the desktop app, GPT-6 Sol still needs a paid plan, and the GPT-5.6 models stay selectable during the rollout 15 14.
• Plus, at $20 a month: no published weekly credit count, and the third-party guides that once quoted community-observed weekly figures now decline to reconstruct one at all, warning that token counts and subscription allowances measure different things and pointing readers at the usage dashboard 5. Plus now includes GPT-6 Astra in Work and Codex following the staged September rollout 17, and since 22 September it also carries GPT-6 Sol and GPT-6 Luna in Work and Codex at half the GPT-5.6 family's credit rate, with Codex lead Tibo Sottiaux saying OpenAI is raising the Plus baseline at the same time 15 13. Eligible Plus accounts can buy credits after limits 18, and since early September they can also buy an instant reset 1.
• Pro, now three tiers. Pro 100 at $100 a month carries 5x higher Codex usage than Plus, and Pro 200 at $200 a month still carries 20x for eligible existing subscribers, a protection OpenAI's help center runs through 29 October 2026 16 13. New Pro 200 subscriptions since the 29 September relaunch buy 10x instead, and from 30 October the $200 plan's Work and Codex allowance is 10x across the board, which Sottiaux says nets out at roughly half the old allowance measured at API list prices 19 20. Pro 500 launched at $500 a month with 25x the Plus allowance and the only Ultrafast mode for GPT-6 Astra, so every tier now costs $20 per unit of Plus allowance 21 22. Pro 100 and Pro 200 include GPT-6 Astra in Chat, Work, and Codex; Astra demand pushed OpenAI to pause new Pro sign-ups on 10 September, and the 29 September relaunch ended that pause 17 13.
• Business: per-seat baseline limits, with shared workspace credits available to extend them 16 23. Business Premium was the first business tier into GPT-6 Astra, on 4 and 5 September 24.
• Enterprise and Edu with flexible pricing: no fixed rate limits, because usage scales with purchased credits 16.

Kinds of reset, and where each one comes from

• Automatic: the rolling seven-day anchor above. Costs nothing, cannot be saved, and arrives whether you want it or not; users have asked for the ability to opt out of surprise resets entirely, so far in vain 1 8.
• Banked: launched with a changelog entry on 11 June 2026 5. The launch announcement granted one free banked reset to Go, Plus, Pro, and Business accounts, but delivery was uneven: GitHub issues document accounts that were never credited, and the changelog's own wording scopes the grant to Plus and Pro 25 5 26 27. A referral promotion ran through 24 June 2026, banking an extra reset for both sides when a referred friend sent their first Codex message, capped at up to three referred friends 28 25. Banked resets expire roughly 30 days after they are granted, and users have reported losing unspent ones to that clock 29 30. Redeeming one refreshes both your five-hour and weekly windows and moves your weekly reset date 9. The system found its strangest job in September: when GPT-6 Astra launched as a restricted preview that almost no paying subscriber could use, Sottiaux announced one banked reset for every day a paid plan spent without Astra, with the first credit landing within about three hours, and Sam Altman apologized for the rollout 4.
• Purchased: new since early September. Eligible Plus and Pro personal accounts can buy an instant reset from Usage settings in the desktop app before reaching a limit, or from an in-app banner after the weekly limit trips; the option is not available on Free, Go, Business, Enterprise, or Edu plans, and availability varies by account and billing country 1. The banner does not appear for a five-hour-only exhaustion, though the settings option can still be there 1. A purchase immediately restores both the five-hour and weekly usage, pulls your normal weekly allowance forward rather than adding a separate entitlement, and restarts your seven-day anchor at your next request; refunds are generally not offered, and the price is shown at checkout rather than published 1. The option surfaced in August testing as an $8 button offered to some Plus accounts once their weekly allowance ran dry, which an OpenAI spokesperson described as exploring ways for capped users to buy more usage 31.

The banked system replaced an older habit: OpenAI used to reset weekly limits for everyone at once to celebrate milestones or apologize for bugs, and moved away from blanket resets as the default practice after users who had planned their week around their limits pushed back 28. Blanket resets did not disappear: all paid plans received one on 9 August 2026 32, and on 21 August 2026 OpenAI announced the 20-million-active-users milestone and credited every paid Codex and ChatGPT Work user with a banked reset voucher, redeemable at a time of the user's choosing within the roughly 30-day bank window 33 34.

The math: what a reset is worth, and the cheaper path

Since OpenAI prices Codex in credits per million tokens, the value of any reset is a burn-rate question, and the burn rates are published. On the current rate card, GPT-5.6 Sol costs 100 credits per million input tokens and 500 per million output, on promotional pricing that runs at least through 21 November 2026; GPT-5.6 Terra runs 50 and 300; GPT-5.6 Luna runs 5 and 30; GPT-6 Astra runs 250 and 1,250; and GPT-5.5 held at 125 and 750 until its retirement from Chat, Work, and Codex on 14 October 2026 35 36 14. On top of those, 22 September added GPT-6 Sol at 50 credits per million input and 250 per million output and GPT-6 Luna at 2.5 and 12.5, exactly half the GPT-5.6 pair's credit rate 15 14. In plain terms, 100 credits buys a million input tokens or 200,000 output tokens on GPT-5.6 Sol, twenty million input tokens on GPT-5.6 Luna, and two million input tokens on GPT-6 Sol, while the Astra ladder burns 2.5 times GPT-5.6 Sol's rate and 5 times GPT-6 Sol's rate in both directions 35 15. The card's reference model is now its newest: a typical Codex task on GPT-6.1 Sol, which began rolling out across eligible paid accounts including Pro in the 29 September announcements, may consume 2 to 15 credits, and Fast mode multiplies the bill when enabled, printed at 2x the standard rate on the card while the GPT-6 launch tables put those models at 2.5x 35 37 15. The card has also put average spend at $100 to $200 per developer per month 35. What the company still does not publish is the weekly credit count for individual plans, so the honest way to price a reset is against your own dashboard: a reset refills whatever your plan's week holds, at whatever model you point it at 5.

The Pro ladder used to reward concentration, and for eligible existing subscribers it still does until 29 October: going from $100 to $200 doubles the subscription but buys four times the usage, cutting the per-unit cost in half 16. The 29 September relaunch ends that discount: new Pro 200 subscriptions buy 10x rather than 20x, the remaining 20x allowances convert on 30 October, and every tier now costs $20 per unit of Plus allowance, with Pro 500 at 25x as the new top rung 19 21 13. The cheapest lever inside any plan is the model, and the lever has moved twice since late August: GPT-5.4 and GPT-5.4 mini left Codex for ChatGPT sign-ins on 31 August 2026, and then GPT-6 Sol and GPT-6 Luna arrived on 22 September at half the GPT-5.6 pair's rate, so the routine-work habit of dropping to a cheaper model now points at GPT-6 Luna, whose output rate of 12.5 credits per million undercuts GPT-6 Sol's 250 by twenty times; there is no GPT-6 Terra, the GPT-5.6 family stays selectable during the phase-out, and GPT-5.5's 14 October exit pointed paid plans at GPT-6 Sol 35 9 15 14 36.

When the cap bites, the fallback differs by plan type. Eligible Plus and Pro users can buy credits, which draw after included usage 18. The other exit is an API key: subscription caps do not apply to API-key sessions, which bill at standard platform rates, with GPT-6 Astra on the API at $10 per million input tokens and $50 per million output and GPT-6 Sol at $2 and $10 after the September halving 16 24 15. If you drain the weekly cap regularly, that is the signal to run heavy agent loops on an API key and keep the subscription for interactive work; an API connection is a separate billing path and does not refill your subscription allowance 5.

What changed in 2026, date by date

• 2 April: pricing moves from per-message to token-based credits for new and existing Plus, Pro, and Business, and new Enterprise plans 16 38.
• 9 April: a limit-system update triggers a wave of community threads recalculating what their plans buy 39.
• 23 April: token pricing extends to remaining Enterprise, Edu, Health, Gov, and Teachers plans 16.
• 11 June: reset banking launches, and the announcement grants one free banked reset to Go, Plus, Pro, and Business; GitHub issues document accounts that never received it 5 25 26.
• 18 June: a GitHub issue asks OpenAI to extend the 30-day banked-reset expiry 29.
• 24 June: the referral-reset promotion ends 28.
• 12 July: the rolling five-hour window is suspended for Plus, Business, and Pro accounts, alongside a one-time usage reset for everyone and changes that made GPT-5.6 Sol consume less allowance 11 12.
• 9 August: a team-wide reset lands for all paid plans 32.
• 21 August: every paid Codex and ChatGPT Work user receives a banked reset voucher celebrating 20 million active users 33 34.
• 25 August: the five-hour window returns for ChatGPT Plus on Codex and Work, so the weekly cap is no longer the only meter on that plan 2 3.
• 26 August: o3 retires from ChatGPT; the retirement does not touch API access 35.
• 31 August: GPT-5.4 and GPT-5.4 mini retire from Codex for ChatGPT sign-ins, with GPT-5.6 Terra and GPT-5.6 Luna named as the replacements; API-key sessions are unaffected 35 9.
• 3 September: GPT-6 Astra launches as a restricted preview almost no paying subscriber can use, and Sottiaux promises one banked reset for every day a paid plan waits without Astra; Sam Altman apologizes for the rollout 4.
• 4 to 7 September: Astra reaches Pro, Enterprise, and Business Premium in Work and Codex and goes live on the API at $10 per million input and $50 per million output, then reaches Plus in Work and Codex; demand pushes OpenAI to pause new Pro sign-ups 24 17.
• Early September: the help center documents purchasable instant resets for eligible Plus and Pro accounts, which August testing had shown as an $8 button for some Plus users 1 31.
• 22 September: GPT-6 Sol and GPT-6 Luna launch in ChatGPT Work and Codex for paid plans at half the GPT-5.6 pair's credit rate, with GPT-6 Luna also reaching Free and Go in the desktop app 15 14.
• 29 to 30 September: in announcements around DevDay 2026, OpenAI reopens Pro 200 at a 10x allowance for new buyers with existing subscribers protected at 20x through 29 October, launches Pro 500 at $500 a month with 25x and the Ultrafast Astra mode, and starts rolling out GPT-6.1 Sol to eligible paid accounts 13 19 21 37.
• 14 October: GPT-5.5 officially retires from ChatGPT, ChatGPT Work, and Codex across all plans; it remains available via the OpenAI API Platform 36 40.

The pattern across those dates: limits got tighter and more metered while resets got more user-controlled, and September added the last step, from user-controlled to purchasable, then closed the month by shrinking what the top subscription buys. Your move, if the weekly cap is a recurring wall, is to buy concentration (Pro 500 at 25x, since the $200 tier drops to 10x from 30 October), efficiency (GPT-6 Sol or Luna for routine work), escape (an API key for the heavy loops), or simply a reset.

Verified October 2026 against OpenAI's help-center articles, community threads, GitHub issues on the openai/codex repository, and third-party documentation; the numbers above change, and this page is maintained rather than published-and-forgotten.