← back to the corkboard
Napkin math · October 17, 2026

What a Reset Is Worth

Tibo's 28-day challenge promises a clear improvement or a full reset of Codex limits every day. A reset is worth the share of your weekly limit you have used minus the share of your week that has passed, and at my assumptions one press of the button costs OpenAI about $11M of compute. Every assumption is adjustable.

On October 4, 2026, Tibo (Thibault Sottiaux), who works on Codex and ChatGPT at OpenAI, announced a 28-day challenge that began the next day: each day OpenAI either ships "one thing that is a clear improvement and relevant for most codex/work users" or ships "a full reset" of usage limits (Sottiaux 2026a). Through Day 13 there have been five global resets, which OpenAI applies to every paid account at once, and one banked reset, which each account holds and applies when it chooses. Tibo calls the global reset his reset button (Sottiaux 2026h), so this post counts one global reset as one press. It works out what a reset is worth to a subscriber, and what one press costs OpenAI.

The short version:

  1. Measured in weeks of your plan, a reset is worth the share of your weekly limit you have used minus the share of your week that has passed. It is worth that to someone who runs out of usage; to someone who never reaches the limit, it is worth nothing.
  2. With a reset every 2.68 days, the pace of the challenge so far, a weekly limit acts as a 2.68-day limit. Someone who would use more than their limit can use up to 2.6 times as much as without resets.
  3. One press costs OpenAI the extra usage of the accounts that would have run out. At my assumptions that is about $11M of compute, most of it on Pro plans, within a range of $0.4M to $270M.

Max Woolf priced a reset on a $100 plan at $25, a quarter of the monthly price, if you use your whole quota (Woolf 2026). A forum user and an essay by WISE AI worked out that a reset also moves your weekly refill later (OpenAI Developer Community 2026b; WISE AI 2026). Two earlier cost estimates started from the prices OpenAI was reported to be testing for a paid reset: TestingCatalog put each free reset at "at least $15-20 million" with about 3M weekly users (TestingCatalog 2026), and Peter Gostev put the August 21 banked reset at $150-240M with 20M users, adding that the real figure is lower because not everyone drains their limits (Gostev 2026). A price measures the revenue OpenAI would give up if every user would otherwise have bought a reset, a point several replies to Gostev made. WISE AI counted the extra usage a reset causes, for a hypothetical 1,000 accounts. This post counts that extra usage across all paid accounts, and you can change every assumption in the interactive figures.

The challenge so far

The 28 days · Oct 5 - Nov 1, 2026
5 global resets · 1 banked reset · 13 days so far
global banked shipped
Day 13 · Sat, Oct 17
Every paid account received a global reset.
landed 2:12 PM PT
Tibo's post ↗
Choose a day to read what shipped or when the reset landed.

Day 2 shipped four things and then reset after a poll in which 76% of 74,565 votes asked for a reset (Sottiaux 2026b; Sottiaux 2026c). Day 3 shipped GPT-6 in ChatGPT and loaded a banked reset into "everyone's paid accounts" (Sottiaux 2026d). The global resets during the challenge have landed 2.68 days apart on average, and two more landed just before it, on September 26 and October 2 (codex-resets.com 2026). That is about twice the pace from August 1 to October 4, when 13 global resets came about 5 days apart, and slower than a 20-day stretch in July that had 11.

How a reset works

Plus and Pro plans have a weekly usage limit, which Codex shows as a percentage used or remaining, with a reset time. Plans are sized as multiples of Plus: Pro 100 is 5x, and Pro 200 is 10x, or 20x through October 29 for subscribers from before September 29 (Sottiaux 2026e; OpenAI 2026c). A third party reports Pro 500 at about 25x, a figure OpenAI has not published (InventiveHQ 2026). When your week ends, your usage returns to zero and whatever you did not use is gone. This post calls that scheduled return to zero your refill, and uses "reset" only for the extra ones OpenAI hands out.

A global reset sets your usage to zero now and starts a new week from your next request, so your next refill comes 7 days after that (OpenAI Developer Community 2026a). OpenAI's help pages describe the same restart for resets you buy and for banked resets (OpenAI 2026a; OpenAI 2026b). OpenAI support wrote in July that "a forced reset can reduce the value of a later natural reset for people who carefully plan around it. That's one of the reasons banked resets were introduced" (OpenAI Developer Community 2026b).

One reset, one subscriber

The rule

A reset gives you back the share of your limit you have used. It also moves your next refill from its old date to 7 days from now. That is as many days later as have already passed in your week: if your week started 5 days ago, your refill was due in 2 days and is now due in 7. Your limit refills once a week, so 5 more days of waiting costs 5/7 of a week of usage, which is the share of the week that has passed.

value of a reset = share of your limit used - share of your week gone

The value is in weeks of your plan. On Pro 100, one week is $23.08 at the plan's price and about $330 at API list prices, which is what the same tokens would cost bought through OpenAI's API.

There is a second way to see the same number, which counts the reset itself as a refill. Your refills then come today, in 7 days, and in 14 days, instead of in 2, 9, and 16, so the whole schedule moves 2 days earlier, which is worth 2/7 of a week. Against that, you lose the part of your limit you had not used. If your refill was due tomorrow, a reset moves your schedule up by a day, worth one seventh of a week, and costs you whatever you had left.

Examples

All rows are for Pro 100.

UsedRefill was due inWeek goneValueAt $23.08 a week
100%1 day86%+0.14 weeks+$3.30
80%4 days43%+0.37 weeks+$8.57
50%3.5 days50%0$0
30%2 days71%-0.41 weeks-$9.56
100%6 days14%+0.86 weeks+$19.78

When you have used your whole limit and your refill is due the next day, a reset is worth one seventh of a week, $3.30 on a $100 plan. If you spend at a pace that would use exactly your limit by the end of the week, a reset is worth nothing. When you have used less of your limit than of your week, and you were saving the rest for the days before your refill, a reset costs you the difference. That is the complaint in replies to Tibo's reset posts such as "So I gain 1% usage and the next reset date pushes 2 days later" (murica1stmuslim 2026).

When a reset is worth nothing

The rule counts the usage you are allowed, whether or not you would spend it. If you do not reach your limit this week or in the weeks after it, the usage a reset gives back and the usage it takes away are both usage you would not have spent, and the reset is worth nothing to you. OpenAI has not published how many subscribers reach their weekly limit. The closest public figure is from Anthropic, which estimated that its Claude Code weekly limits would affect fewer than 5% of its subscribers (TechCrunch 2025).

Your reset
Your plan
Pro 200 is 20x until Oct 29 for subscribers from before Sep 29, 10x otherwise. Pro 500's 25x is reported, not official.
During the challenge, count from the last reset: 2.68 days after one, your refill was due in about 4.3 days.
You get back+80% of a week
Your next refill is pushed back by 3 days-43% of a week
+37% of a week= $8.57 at what you pay ($23.08 a week)about $123 at API list prices
80% used - 43% of the week gone = +37%
This counts the usage you are allowed. It is worth this only if you would have hit your limit before your next refill; otherwise it is worth $0 to you.
A reset gives back what you used and restarts your seven-day clock.

Resets that keep coming

With a reset every 2.68 days, your week ends at the next reset before its 7 days are up. Since September 26, the longest gap between global resets has been about 6.1 days. During a run of resets, count the share of the week from the last reset, since each reset restarts the clock. For someone who uses up their limit between every pair of resets, a reset g days after the previous one is worth 1 - g/7 weeks, which at the average gap of 2.68 days is 0.62 weeks.

When resets come every k days, with k under 7, the most you can use in a week is 7/k weekly limits, which at the challenge's pace is 2.6. If the pace holds for all four weeks, a subscriber who would use three times their limit gets about 10.4 weeks of usage over the challenge instead of 4.

Your plan with resets every few days
Your plan
Pro 200 is 20x until Oct 29 for subscribers from before Sep 29, 10x otherwise. Pro 500's 25x is reported, not official.
without resets
without resets: 1.0 weeks of quota per week
with resets
with resets: 2.0 weeks of quota per week
weeks of quota a week
Your plan acts like 2.0x its sizeabout $660 a week at API prices instead of $330
With a reset every 2.68 days, your weekly limit refills every 2.68 days.

Averaged over subscribers

For the cost to OpenAI, what matters is what a reset is worth on average to the accounts it reaches. Take a subscriber who would use D times their weekly limit if there were no limit, spending at an even pace until they run out. Three cases cover the resets that occur:

  1. A surprise reset, landing at a random point in each subscriber's week, is worth (1 - 1/D)/2 weeks on average. That is under half a week even for the heaviest subscriber, because a random reset lands, on average, halfway through the week.
  2. A banked reset, applied the moment the subscriber hits the limit, is worth 1 - 1/D, twice as much. When no global reset is about to land, that moment is the best time to use one.
  3. A reset at the challenge's pace, 2.68 days after the previous one, finds an account that resumed right after that previous reset 38% of the way through its week. It is worth min(1, 0.38 D) - 0.38.
Would use, with no limitSurpriseBankedChallenge pace
1.5x the limit0.17 weeks0.33 weeks0.19 weeks
2x0.250.500.38
3x0.330.670.62
5x0.400.800.62

Below 1x, all three are zero. Accounts that run out do not all go over by the same amount, and the cost estimate assumes many go a little over and a few go far over. With an average of 2x, a press at the challenge's pace is worth 0.31 weeks per account that runs out, a surprise reset 0.20, and a banked reset 0.40. These are below the 2x row of the table because accounts just over their limit get little from a reset, and accounts far over it gain no more than the cap.

One press of the button

The formula

One press costs OpenAI the extra usage it unlocks, priced at what that usage costs to serve:

  1. Take the paid accounts that used Codex or ChatGPT Work that week, split by plan.
  2. Keep the share on each plan that would run out before their next refill; the rest cost nothing.
  3. Multiply by the extra usage per such account from the previous section, in weeks of their plan.
  4. Multiply by what one week of that plan is worth at API list prices.
  5. Multiply by the cost to serve, as a share of API list price.

The inputs

Each input has a default and the range I consider plausible; the ones without a source are my assumptions.

  1. Paid accounts using Codex or Work in a week: 12M, range 6-24M. OpenAI reported 40M active users across Codex and Work on October 7 (Sottiaux 2026d). Codex is also in the Free and Go plans (OpenAI 2026d), which get no reset. 12M assumes 30% of the 40M are paid.
  2. Share of those on Pro plans: 10%, range 5-15%. Of the Pro accounts, Pro 100 is half, Pro 200 40%, and Pro 500 10%. Every Pro 200 account is counted at the old 20x allowance, which only subscribers from before September 29 keep through October 29.
  3. Plus accounts that run out: 7%, range 4-15%. Plus here includes Business Standard and Enterprise seats without flexible pricing, which have Plus-level limits. Anthropic estimated under 5% for its weekly limits (TechCrunch 2025) and about 7% for a 2026 session-limit change (The Register 2026).
  4. Pro accounts that run out: 30%, range 15-50%. Pro is bought for heavy use.
  5. Average use with no limit, among accounts that run out: 2x the limit, range 1.5-3x. There is no public data.
  6. One Plus week at API list prices: $66, range $55-81. FinOps LLM measured a median of $67 (FinOps LLM 2026a), and InventiveHQ projected $66 from its Pro measurements (InventiveHQ 2026). Pro tiers use their advertised multiples. FinOps LLM measured a median Pro 100 week at 2.6x a Plus week (FinOps LLM 2026b); if every Pro tier fell short of its multiple by that ratio, the Pro part of the cost would be about half.
  7. Cost to serve, as a share of API list price: 10%, range 5-30%. SemiAnalysis assumes a 92% API gross margin in its model of Anthropic's plans (SemiAnalysis 2026), and Alderson puts compute at about 10% of Anthropic's API prices (Alderson 2026). For OpenAI, The Information reported a 70% compute margin on paying users as of October 2025 (Fortune 2025), so compute was about 30% of what they paid.

The result

One press of the button
Cost assumptions
6M-24M · my assumption
5-15% · my assumption
4-15% · near Anthropic's 5-7%
15-50% · my assumption
1.5-3x · no public data
FinOps LLM measured a median of $67; InventiveHQ projected $66.
5-30% · SemiAnalysis 8%, Alderson ~10%
When the reset lands
Pro 200 allowance
Pro tiers are priced at their advertised multiples of a Plus week.
ONE PRESS: $11.0M of compute· $110M at list
THE RESET BUTTON
ONE PRESS
OCT 17, 2026
NAPKIN MATH
ACCOUNTS THAT RUN OUT x EXTRA WEEKS x COST OF A WEEK
PLUS
10.8M accounts x 7% run out
756K x 0.31 wk x $6.60
COST$1.53M
PRO 100
600K accounts x 30% run out
180K x 0.31 wk x $33
COST$1.82M
PRO 200
480K accounts x 30% run out
144K x 0.31 wk x $132
COST$5.83M
PRO 500 (REPORTED)
120K accounts x 30% run out
36K x 0.31 wk x $165
COST$1.82M
COMPUTE COST
$11.0M
AT API LIST PRICE
$110M
PER PAID ACCOUNT
$0.92
THE CHALLENGE SO FAR · 5 PRESSES
COMPUTE COST$51.2M
*** see you in about 2.68 days ***
What one press costs OpenAI: the extra usage of accounts that would run out, at the cost to serve. Change any number; the link keeps them.

At the defaults, one press at the challenge's pace costs about $11M of compute, $110M at API list prices, or about $0.92 per paid account in use that week. Pro accounts are 10% of the accounts and 86% of the cost, and Pro 200 on the old 20x allowance is about half of the total. The five global presses through Day 13, each priced at its own gap from the previous reset, come to about $51M of compute. A surprise press costs about $7.2M. A banked reset like the one on Day 3 costs about $14.5M if every account that runs out applies it the moment it hits the limit.

Moving one input across its range while the others stay at their defaults changes the result 6x for the cost to serve, 4x for the number of paid accounts, and 2.8x for the share of Pro accounts that run out. The share of Plus accounts that run out changes it by about a fifth. With every input at the low end, a press costs about $0.4M; with every input at the high end, about $270M.

What this leaves out

Capacity. If OpenAI's GPUs are bought ahead of demand, the extra tokens a press serves add little cost, and the larger cost may be the peak load it creates. A global reset refills every paid account within a few hours, and the accounts that had run out come back at once. Tibo has described a 5-hour limit on Pro as a way "to smoothen the demand curve" (Sottiaux 2026f), and Day 11 added compute and changes to cut "Selected model is at capacity" errors (Sottiaux 2026i). When OpenAI has no spare capacity, a press costs what the same compute would have earned on the API, which is the API list price line on the receipt.

Behavior. The model assumes people spend at an even pace. Some replies to Tibo's posts describe using up the remaining quota between an announcement and the moment the reset lands, which makes a press cost more than the model says (Aoakudotcom 2026).

The value to OpenAI. Most of a press goes to the heaviest accounts, and the replies under the challenge posts include heavy users who say they are moving to Claude, such as one moving "All 3 Pro x20 plans and the $1000/month API bill" to Anthropic (Entinode 2026). The share of these users a press keeps would show whether a press is worth its cost, and OpenAI has not published it.

What a week is worth. Day 2 stopped counting auto-review against usage, which Tibo put at 2-10% of a plan for those who use it (Sottiaux 2026g), and Day 12 used caching to make the same quota cover more requests (Sottiaux 2026j). Both make a week of quota cover more work, and so make a week worth more at API list prices. After October 29, Pro 200 subscribers from before September 29 move to the lower allowance (OpenAI 2026c), which lowers the cost of a press by about a quarter at the defaults; the receipt has a switch for it.

The inputs OpenAI has. The paid count, the share that runs out on each plan, the spread of overshoot, and the cost to serve are in OpenAI's own data; here they are my assumptions. If you have the real numbers, the figures take them, and the copy-link button shares them.

For a subscriber, two rules follow. A reset helps you when you have used more of your limit than of your week. A banked reset is worth the most when you apply it the moment you hit the limit, on a day when no global reset is about to land.

References

  1. Alderson, M. (2026). "No, it doesn't cost Anthropic $5k per Claude Code user." March 9, 2026. https://martinalderson.com/posts/no-it-doesnt-cost-anthropic-5k-per-claude-code-user/
  2. Aoakudotcom (2026). Reply on X, October 10, 2026. https://x.com/Aoakudotcom/status/2109059005297361057
  3. codex-resets.com (2026). Codex reset tracker. Accessed October 17, 2026. https://codex-resets.com/
  4. Entinode (2026). Reply on X, October 4, 2026. https://x.com/Entinode/status/2106845900748558408
  5. FinOps LLM (2026a). Measurements of the Codex weekly limit on ChatGPT Plus. Updated October 10, 2026. https://finopsllm.com/research/codex-usage-limits-chatgpt-pro
  6. FinOps LLM (2026b). Measurements of six Pro 100 weekly windows. October 9, 2026. https://finopsllm.com/research/chatgpt-pro-100-real-codex-multiplier
  7. Fortune (2025). Report on The Information's figures for OpenAI's compute margins. December 21, 2025. https://fortune.com/2025/12/21/openai-compute-margins-revenue-chatgpt-earnings-sam-altman/
  8. Gostev, P. (2026). Post on X, August 21, 2026. https://x.com/petergostev/status/2090817766051762189
  9. InventiveHQ (2026). Measurements of Codex subscription usage. October 4, 2026. https://inventivehq.com/blog/codex-subscription-usage-measured
  10. murica1stmuslim (2026). Reply on X, October 12, 2026. https://x.com/murica1stmuslim/status/2109817607347949772
  11. OpenAI (2026a). "Paid weekly Work and Codex rate limit resets." Help Center, accessed October 17, 2026. https://help.openai.com/en/articles/20001507-paid-weekly-work-and-codex-rate-limit-resets
  12. OpenAI (2026b). "How banked Codex resets work." Help Center, accessed October 17, 2026. https://help.openai.com/en/articles/20001498-how-banked-codex-resets-work
  13. OpenAI (2026c). "About ChatGPT Pro tiers." Help Center, accessed October 17, 2026. https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
  14. OpenAI (2026d). Codex pricing. Accessed October 17, 2026. https://learn.chatgpt.com/docs/pricing
  15. OpenAI Developer Community (2026a). OpenAI Support reply on a moved weekly reset date, October 6, 2026. https://community.openai.com/t/my-end-of-week-date-got-moved/1396556
  16. OpenAI Developer Community (2026b). Thread on usage lost to resets, July 2026. https://community.openai.com/t/this-is-ridiculous-im-losing-usage-due-to-the-resets/1388195
  17. SemiAnalysis (2026). Analysis of AI subscription value per dollar. October 5, 2026. https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x
  18. Sottiaux, T. (2026a). Post on X announcing the challenge, October 4, 2026. https://x.com/thsottiaux/status/2106845241357824205
  19. Sottiaux, T. (2026b). Day 2 poll, October 6, 2026. https://x.com/thsottiaux/status/2107576143285219799
  20. Sottiaux, T. (2026c). Day 2 reset, October 6, 2026. https://x.com/thsottiaux/status/2107676072871600470
  21. Sottiaux, T. (2026d). Day 3, October 7, 2026. https://x.com/thsottiaux/status/2107913674593644711
  22. Sottiaux, T. (2026e). Plan multiples, September 29, 2026. https://x.com/thsottiaux/status/2104951965184925941
  23. Sottiaux, T. (2026f). On the Pro 5-hour limit poll, October 11, 2026. https://x.com/thsottiaux/status/2109439315570999598
  24. Sottiaux, T. (2026g). Day 2 roundup, October 6, 2026. https://x.com/thsottiaux/status/2107575657014468879
  25. Sottiaux, T. (2026h). Day 10 reset, October 14, 2026. https://x.com/thsottiaux/status/2110595484180214059
  26. Sottiaux, T. (2026i). Day 11, October 15, 2026. https://x.com/thsottiaux/status/2110750421367287996
  27. Sottiaux, T. (2026j). Day 12, October 16, 2026. https://x.com/thsottiaux/status/2111244596131111253
  28. TechCrunch (2025). "Anthropic unveils new rate limits to curb Claude Code power users." July 28, 2025. https://techcrunch.com/2025/07/28/anthropic-unveils-new-rate-limits-to-curb-claude-code-power-users/
  29. TestingCatalog (2026). Post on X, August 7, 2026. https://x.com/testingcatalog/status/2085653957997682695
  30. The Register (2026). Report on Anthropic's session-limit change. March 26, 2026. https://www.theregister.com/software/2026/03/26/anthropic-tweaks-claude-usage-limits-to-manage-capacity/5225406
  31. WISE AI (2026). "How much does Tibo give away with one Reset?" September 14, 2026. https://www.spke.com/en/article/tibo-reset-ai-cost
  32. Woolf, M. (2026). "What's the deal with all the random weekly quota resets for agents lately?" July 18, 2026. https://minimaxir.com/2026/07/agent-quota-reset/