When Codex quotas get a surprise reset, this page catches it — signals attached.

Codex

What this word means

Watching — nothing yet

The radar is awake and reading its sources on schedule; there is simply nothing worth interrupting you for.

A watch opens when archived evidence clears the scoring bar.

Every word we publish

From the record

48-hour base rate

Measured, not modelled. Slide a 48-hour window across every day we have on record, one hour at a time, and count how many of those windows contained a confirmed reset. That is the 44%.

It says what history does, not what today’s evidence says. The moment evidence clears our bar, a reading of that evidence takes this number’s place.

Resets on record
29
Days covered
99.3 days
Windows that held one
1,016 of 2,335
Every reset on the record

Day 5 of this gap. The median is 2.2 days, and 21 of the 28 gaps on record had closed by now.

Chosen by rule, not written by a model

What to do

Every reading carries one action from a fixed list of six, picked by rule from the state and the evidence — the bold line above is that action in plain words. A page that has gone stale always falls back to no action.

Continue normally.
now
Prepare the next heavy task; no need to change plans yet.
If you already planned work, use remaining capacity before a possible refresh.
Wait for the reset to land before starting quota-heavy work.
Reset landed; if your account reflects it, start heavy work.
No action recommended.
No signal. Spend freely. Get pinged the moment it's confirmed

Why the number moved — last 7 days

  • 05 Sep93%Tibo: “we will do the full banked reset today too for all Plus, Pro and Business users. Lands end of day.”
  • 04 Sep92%Tibo: some Plus and Business users miss Astra today, so a banked reset "lands by end of day."
  • 03 Sep18%Folks are burning Codex limits on faith that Astra will bring a reset. That is hope, not a live grant.
  • 02 Sep52%Fresh reports say the new weekly limit is draining far faster than expected.
  • 01 Sep22%Sunday’s paid-usage reset already landed. Fresh posts are jokes, leftover burn talk, and wishes for another Tibo drop.
  • 31 Aug58%OpenAI has identified a major quota issue, keeping another recovery reset in play.
  • 30 Aug97%Tibo says the Codex reset “will land at 6pm PST.”

Nothing brewing right now.

No signal on show — nothing is being claimed.

How old this page is

Freshness

The radar reads its sources every five minutes and the site republishes at least every fifteen. An unchanged cycle deliberately skips its redeploy, so a perfectly healthy page can be a little older than its last check.

Rather than pretend otherwise, the page grades its own age and withdraws what it can no longer stand behind.

Live
under 15 min
Aging
15-30 min
Stale, number withdrawn
past 30 min
Outage, said out loud
past 60 min
Every word we publish

Signal desk

What signals the number, right now.

Six weeks

The last six weeks of Codex, as recorded.

Last reset

31 Aug 02:29 UTC

The newest verified reset

The newest reset verified from the vendor’s own announcement, independently of what we posted.

What we alerted at the time: “Codex usage has been reset for all paid plans.”

Every reset on the record

On the record

The record

What we hold

Every reset we have verified, each with at least one public signal attached at the moment the row was written — not at the moment it is rendered.

Rows are append-only. A correction is a new row, never an edit, so a link you cite will not quietly change meaning.

Confirmed resets
29
Signals on file
30
First on record
23 May 20:14 UTC
Read the record

Codex

28 Jul

Reset confirmed

CODEX RESET again!

29 Jul

Reset confirmed

Tibo reset usage limits for all ChatGPT Work and Codex users.

01 Aug

Reset confirmed

Global reset, verified from the vendor’s own announcement. Our alert went out on 01 Aug, 2h later.

08 Aug

Reset confirmed

Global reset, verified from the vendor’s own announcement. Our alert went out on 09 Aug, 8h later.

09 Aug

Reset confirmed

Tibo just reset Codex limits for all paid users.

11 Aug

Reset confirmed

Reset landed for all paid ChatGPT Work and Codex users!

13 Aug

Reset confirmed

CONFIRMED — the Codex reset has landed.

24 Aug

Reset confirmed

The reset has landed and has been propagated to accounts.

25 Aug

Reset confirmed

RESET LANDED — CODEX

27 Aug

Reset confirmed

Codex usage just got a fresh global reset.

30 Aug

Reset confirmed

Codex reset landed for all paid users!

31 Aug

Reset confirmed

Codex usage has been reset for all paid plans.

The next one lands while you are not looking.

Get the next alert

Covers Codex. Checks every five minutes; the ping goes out the moment a reset is verified — high-confidence watches, landings, retractions, nothing else.

Track record

29 confirmed Codex resets on record. Every one with its signals on file.

History, not a forecast.

Measured from the record

From the record

Median gap

Half the gaps between confirmed resets were shorter than this and half were longer. It is the middle of the record, not an average, so one very long drought cannot drag it.

Every reset on the record
2.2d

From the record

Longest drought

The longest stretch on record between two confirmed resets. It is the honest ceiling on any expectation you form from the median.

Every reset on the record
11d

From the record

Confirmed resets

How many resets we have verified and recorded. Every one of them is a row in the record with its own signals on file.

Every reset on the record
29

From the record

Signals on file

Public sources attached to those rows. More than one per reset means some were vouched for twice.

Every reset on the record
30
  1. 3.4d
  2. 3.7d
  3. 11d
  4. 2d
  5. 2.2d
  6. 7.7d
  7. 3d
  8. 1d
  9. 2.3d
  10. 4.1d

From the record

The gaps between resets

Each bar is one stretch with no confirmed reset, newest first. The dotted line is the median.

Both ends of every gap are in the record below, with their signals.

Every reset on the record

  1. 00
  2. 01
  3. 02
  4. 03
  5. 04
  6. 05
  7. 06
  8. 07
  9. 08
  10. 09
  11. 10
  12. 11
  13. 12
  14. 13
  15. 14
  16. 15
  17. 16
  18. 17
  19. 18
  20. 19
  21. 20
  22. 21
  23. 22
  24. 23

From the record

Confirmed resets by hour, UTC

One column per hour of the UTC day, across all 29 confirmed resets on record. The tallest column is 00:00 UTC with 5.

It is a shape, not a schedule. Nothing in our alerting reads this chart, and a vendor is free to break the pattern the day after you notice it.

Every reset on the record

Permanent event log

Every alert on the record — open one to see its signals.

  1. Watch A major quota incident puts another Codex reset in play. 2 signals
  2. Reset confirmed Codex usage has been reset for all paid plans. 2 signals
  3. Reset confirmed Codex reset landed for all paid users! 3 signals
  4. Reset confirmed Codex usage just got a fresh global reset. 3 signals
  5. Reset confirmed RESET LANDED — CODEX 2 signals
  6. Reset confirmed The reset has landed and has been propagated to accounts. 2 signals

The whole of it

Everything behind the numbers above, one fold away.

How resets actually work

Two clocks, not one

Codex subscriptions meter you on two clocks at once. A short rolling window — five hours — governs a burst of work, and a longer weekly ceiling governs the sum of it. Emptying one does not refill the other, which is why work can stop while a counter you were watching still looks healthy. A vendor announcement rarely says which clock it flushed; when it does, the alert carries it.

A reset is not a rollover

Your own five-hour window expiring is a rollover. It happens on your schedule, no announcement exists for it, and no radar can tell you about it — your client can. An unscheduled reset is the vendor flushing counters early, for many people at once. That is the only thing this site watches.

The kinds we archive

Global reset
The vendor flushed counters for everybody on a plan.
Banked-reset grant
A grant handed out on top of the normal cycle, usually as goodwill after something went wrong.
Partial cohort reset
Only some accounts were flushed. The scope says which.
Policy change
The limits themselves moved. Nobody's counter was reset.
Outage recovery
Counters were restored after an incident ate them.
What every word and number means

The states

Watching — nothing yet
The radar is awake and reading its sources on schedule; there is simply nothing worth interrupting you for. A watch opens when archived evidence clears the scoring bar.
Watch
Evidence is building toward a reset and has cleared our scoring bar. It is a reading on public signals, never a promise. It becomes confirmed only when a reset lands and a signal is archived, never because the percentage got high.
Reset confirmed
A reset was confirmed, and the evidence for it is archived and linked. It becomes landed once the reset is observed applying to accounts.
Reset landed
The reset landed and was observed applying. We never read your account. Check the client you actually use.
Watch closed
A watch closed without our confirming a reset. When a reset is later verified for that day anyway, the row says so — and the miss counts against us. It stays on the record rather than being deleted.
Retracted
We published an alert and later judged it wrong. Retractions stay on the record with everything else.

The numbers

48h signal confidence
How confident we are that the public signals we have archived point at a reset landing within 48 hours. A watch opens at 25%, and that bar applies to this number only. It is not your personal remaining quota; we never read your account.
48h base rate
What history alone says. It stands in for the live reading on a quiet day and sits beside it on a watch day, where the two will often disagree. Slide a 48-hour window across the whole record and count how many of those windows held a confirmed reset. It is measured, not modelled, it never opens a watch and never decides an alert — so it can sit above the 25% bar while no watch is open. The two numbers measure different things.

What counts as a signal

Official
The vendor's own status page or documentation.
Team
A named member of the vendor's team, posting publicly.
User report
Somebody describing what they saw in their own client.
Own probes
Codex accounts we run ourselves and read directly.

How fresh the page is

Live
Checked within 15 minutes.
Aging
15 to 30 minutes. Normal on a quiet day, because an unchanged cycle deliberately skips its redeploy. Nothing is withdrawn.
Stale
Past 30 minutes. The signal confidence hides and the action falls back to no action.
Outage
Past 60 minutes. The page says so across its full width — if our system ever stops reporting, that banner is the alarm.

The name

Tibo
@thsottiaux — the OpenAI engineer whose posts announce most Codex resets. The site is named for the community plea in his replies whenever quotas run dry: save us, Tibo.
How an alert gets made
  1. CollectOfficial status pages, what the vendors’ teams post publicly, and Codex accounts we run ourselves.
  2. ScoreScope, signal and conflict rules cut most of it. What survives is scored on signals that led to a reset before.
  3. Cross-checkA second automated reader re-reads the evidence against the alert. Disagreement means nothing ships.
  4. PublishThe page refreshes at least every fifteen minutes. Telegram carries the strong alerts, the landings and the retractions.

Confirmation is gated on evidence, never on a high percentage. A ninety-percent watch with no supporting signal still does not get confirmed, and an alert we later judge wrong is retracted in place rather than deleted.

Most signals today are one OpenAI engineer's posts, and you could follow him directly. What this adds: verification, a record that cannot be quietly edited, and a ping the moment a reset is verified — including when the announcement is not his.

How to read the track record

Codex

Measured track record

Graded against the verified record — resets confirmed from vendor announcements, independently of our alerts. A number with no ready dataset prints rather than a guess.

Real resets in the window
12
Watches the radar opened
26
When it opened one, a reset followed
61.5%
Resets it had a watch open for
75.0%
Resets it called before they landed
66.7%
How early the alert cameahead of the event · target 12h ahead
8.37h
False alarmsour ceiling: 3 a week
1.43

The 29 resets above are everything verified since the record began; this card grades only the span it names — the part of the record we have back-tested so far.

8 of those 26 reached the channel: every alert waits for a human to release it, and a draft that is not released in time expires. That is why the record below is shorter than this count.

Measured 18 Jul – 05 Sep 2026 · first alert 21 Jul

Every reset on the record, with signals

Links on each reset row are the vendor’s announcement itself; an alert’s signals are what we could publicly cite at the time we published.

Codex

29 confirmed resets on record. Every one with its signals on file.

Confirmed resets

  1. Global reset x.com@thsottiaux 3.4d after the previous one
  2. Global reset x.com@thsottiaux 3.7d after the previous one
  3. Global reset x.com@thsottiaux 11d after the previous one
  4. Global reset x.com@thsottiaux 2d after the previous one
  5. Global reset x.com@thsottiaux ·1x.com@thsottiaux ·2 2.2d after the previous one
  6. Global reset x.com@thsottiaux 7.7d after the previous one
  7. Global reset x.com@thsottiaux 3d after the previous one
  8. Global reset x.com@thsottiaux 1d after the previous one
  9. Global reset x.com@thsottiaux 2.3d after the previous one
  10. Global reset x.com@thsottiaux 4.1d after the previous one
  11. Global reset x.com@thsottiaux 3.6d after the previous one
  12. Global reset x.com@thsottiaux 2d after the previous one
  13. Global reset x.com@thsottiaux 1.4d after the previous one
  14. Global reset x.com@thsottiaux 1d after the previous one
  15. Banked-reset grant x.com@thsottiaux 0.9d after the previous one
  16. Partial cohort reset Some accounts x.com@thsottiaux 0.1d after the previous one
  17. Global reset x.com@thsottiaux 1.5d after the previous one
  18. Global reset x.com@thsottiaux 0.5d after the previous one
  19. Global reset x.com@thsottiaux 0.9d after the previous one
  20. Global reset x.com@thsottiaux 9.9d after the previous one
  21. Global reset x.com@thsottiaux 1d after the previous one
  22. Global reset x.com@thsottiaux 1.1d after the previous one
  23. Global reset x.com@thsottiaux 0.9d after the previous one
  24. Global reset x.com@thsottiaux 9d after the previous one
  25. Global reset x.com@thsottiaux 6d after the previous one
  26. Banked-reset grant x.com@OpenAI 8d after the previous one
  27. Global reset x.com@thsottiaux 3.4d after the previous one
  28. Global reset x.com@thsottiaux 7.8d after the previous one
  29. Global reset x.com@thsottiaux

The alerts we published at the time

  1. Watch A major quota incident puts another Codex reset in play. 2 signals
  2. Reset confirmed Codex usage has been reset for all paid plans. 2 signals
  3. Reset confirmed Codex reset landed for all paid users! 3 signals
  4. Reset confirmed Codex usage just got a fresh global reset. 3 signals
  5. Reset confirmed RESET LANDED — CODEX 2 signals
  6. Reset confirmed The reset has landed and has been propagated to accounts. 2 signals
  7. Watch A full Codex reset is likely tomorrow. 4 signals
  8. Reset confirmed CONFIRMED — the Codex reset has landed. 2 signals
  9. Reset confirmed Reset landed for all paid ChatGPT Work and Codex users! 1 signal
  10. Reset confirmed Tibo just reset Codex limits for all paid users. 1 signal
  11. Watch closed Codex usage limits have been reset for the weekend. 4 signals
  12. Reset confirmed Tibo reset usage limits for all ChatGPT Work and Codex users. 3 signals
  13. Reset confirmed CODEX RESET again! 1 signal
  14. Watch Codex has a real shot at a fresh global reset after this outage. 2 signals
  15. Reset confirmed Codex paid-user quotas are refreshed now. 1 signal
Every number on this page, as a table
FigureWhereValue
StateCodexWatching — nothing yet
48h base rateCodex44%
48h windows measuredCodex1,016 of 2,335
Days coveredCodex99.3 days
Days into this gapCodex5.2
Confirmed resetsCodex29
Signals on fileCodex30
Median gapCodex2.2d
Longest droughtCodex11d
Past resetsCodex12
Watches openedCodex26
Watch precisionCodex61.5%
Watch recallCodex75.0%
Resets we called before they landedCodex66.7%
Median heads-up timeCodex8.37h
False alerts / weekCodex1.43
Last checkedPage05 Sep 06:09 UTC
Goes stale afterPage30 min
Declares an outage afterPage60 min

A number we cannot measure honestly prints an em dash. It is never a placeholder, and never a zero standing in for unknown. Past resets are the real, verified resets inside the measured window. Watch precision counts watches, not resets: evidence can build again around the same reset and open a second watch, so a handful of successful watches can point at fewer distinct resets. That is why precision and recall move independently.

What we never do
  • We never read your account. There is no login, no token, no extension, and no way for this site to know your remaining quota.
  • No cookies, no email capture, no analytics that identify you. Traffic counting is cookieless and aggregate.
  • Every public reason resolves to archived evidence. The renderer maps evidence to trusted links; the model never emits a URL.
  • An alert we got wrong stays on the record, retracted in place. The track record is graded against resets verified from vendor announcements, independently of our alerts — one fold above this one.