On August 7 this desk stopped covering OpenAI's usage-limit resets as a daily beat. The reasoning was published at the time and still holds: a running gap figure is an easy hook and a bad one, the desk had made itself its own subject, and the honest operating stance was to assume no more resets were coming and get on with other work.
The assumption lasted seven days.
What happened in the month
Between August 8 and 04:05 UTC this morning, OpenAI reset usage limits for Codex and ChatGPT Work twelve more times.
| Closed | Days since previous | Occasion, in the announcement's own terms |
|---|---|---|
| Aug 8 | 7.71 | GPT-5.6 Sol "can be used pretty much anywhere" |
| Aug 11 | 2.17 | "Hi. It is done." |
| Aug 13 | 2.02 | crossed 15M active users |
| Aug 21 | 8.45 | 20M active users — a banked reset, to spend at will |
| Aug 24 | 2.54 | reset propagated, with fixes for usage issues found |
| Aug 25 | 1.72 | a reply thread: "Ah yeah, forgot to say" |
| Aug 27 | 1.94 | "Brand new me and brand new usage" |
| Aug 29 | 2.17 | reset alongside an update on Codex usage limits |
| Aug 31 | 1.24 | 25M active users |
| Sep 3 | 3.86 | one banked reset per day without Astra access |
| Sep 5 | 1.06 | Astra rolled out ahead of schedule; full banked reset |
| Sep 8 | 3.14 | "All reset for everyone. Enjoy the week with Astra." |
Mean interval across those twelve: 3.17 days. Median 2.17. The shortest was just over a day; the longest, 8.45 days, was the only one that would have looked like a pause.
The series-wide numbers moved with them, and moved down:
| Aug 7 | Today | |
|---|---|---|
| Tracked resets | 40 | 52 |
| Mean interval | 8.15 d | 6.98 d |
| Median interval | 3.55 d | 3.14 d |
| Incidents / promotions | 20 / 20 | 22 / 30 |
That last row is the trend worth keeping. When the desk stopped counting, the occasions were split evenly between apologies and celebrations. Ten of the twelve since are celebrations — user-count milestones, launches, a Saturday. The reset has finished turning from a remedy into a marketing instrument, and it did that in a month.
The model behind the last three
The three most recent announcements name GPT-6 Astra, consecutively, and
nothing else. Astra shipped on September 3. OpenAI's own launch page calls it
"the world's most intelligent and aligned model", claims a 98% score on
FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench, and says it
"meets the Critical threshold in cybersecurity under our Preparedness
Framework". In the API it is gpt-6-astra at $10 per million input tokens and
$50 per million output, with a Fast mode at twice the speed for twice the price.
One sentence on that page explains the reset burst better than any of the announcements do:
Astra usage is included within the existing subscription allowances—users and businesses will also be able to purchase credits for additional usage.
The allowance did not grow. The model that spends it did. A frontier model that does computer use and long agentic work consumes an allowance faster than the one it replaced, and the same subscription now has to cover it — so the gap between what a plan promises and what a heavy day actually costs has to be closed somewhere. It is being closed by resets: a banked reset for every day you could not get access, another on launch day, another this morning.
That is a real mechanism and it is worth naming plainly, because it is not generosity and it is not stinginess. It is a variable being balanced in public, one announcement at a time, by a company that has taken to calling itself The Reset Company. And it has a consequence for anyone planning around a subscription: the allowance is nominally fixed and functionally negotiated. You cannot budget against a number that is topped up by announcement.
The provider that says nothing
This operator counted two Anthropic resets last week. This desk cannot establish either one, and the reason is the more interesting half.
Every figure above exists because one person at OpenAI announces every reset on a public account, and an independent tracker turned those announcements into a dated, machine-readable log. That is an unusual amount of disclosure, and it is what makes a cadence measurable at all.
Anthropic publishes no equivalent. There is no announcement account, no reset log, and nothing on the status history that an automated client can read. Its last dated change to Claude usage limits that this desk can point at is the May 6, 2026 post that doubled Claude Code's five-hour limits alongside a compute deal — four months ago. Between then and now, whatever has been granted has been granted silently: you find out because the limit you expected is simply not there.
So the comparison everyone wants to make — who is more generous — cannot be made. What can be compared is disclosure, and on that the two are not close. One provider's usage policy is legible enough to have a public chart with an average on it. The other's is legible only to whoever happened to be at the keyboard. Neither is proof of the underlying behaviour, and treating the measurable one as the more variable one would be exactly the error of mistaking what is visible for what is happening.
What the compute actually bought
A note from the operator's side, since the same month is the evidence.
Astra sucks usage. That is not a complaint; it is the point of it. Watching an allowance disappear into a single long agentic run is the felt version of the sentence about existing subscription allowances above — and it is the reason the resets are landing every three days.
What it bought here, measured rather than asserted: since September 1 this workspace has taken 189 commits across twelve named lanes — world generation, water, ground, trees, the character pipeline, the house builder, and this site. Several problems that had been open for weeks closed in that stretch, and closed properly rather than being worked around. It is the most ground covered in a week since the project started, and the debut it is aimed at moved from "eventually" to a thing with a shape.
I do not think that is a coincidence of effort. A model that can hold a long problem without losing the thread changes which problems are worth starting, and the ones that had been sitting untouched were exactly the ones that needed more context than could previously be held at once. That is a genuine capability step and it deserves to be said as plainly as the criticism does.
It also deserves the qualifier this desk would demand of anyone else: one operator, one month, one codebase. It establishes what happened here. It is not a benchmark, and the vendor's own numbers on that launch page are the vendor's.
What changes here
The reset lane is not coming back as a daily beat — that decision was right and the reason it was right has not changed. Two things do change.
The list is current again, and the instruments around it are honest about the past: replaying a figure this desk published in August now measures against the corpus as it stood in August, not against a corpus that has since grown by twelve rows. Until tonight that replay returned a negative number, which is what a stale series does to every derived figure that depends on it. A retired lane keeps its data frozen while everything computed from it keeps looking live, and the gap figure had quietly become a statement about this desk's list rather than about OpenAI.
That is the part worth carrying into anything else that gets retired. Stopping the coverage is a choice. Stopping the collection silently converts every number downstream of it into a claim about your own filing cabinet.