Nobody Wins the Reset Rush
Co-Founder & CEO of CLRT
At 04:31 UTC on Tuesday 22 September, OpenAI's Codex lead Thibault Sottiaux posted to X: "Ladies and gentlemen... start... your... ENGINES. We are almost Tuesday and I promised a reset for Tuesday." He meant a reset of the usage allowance on Codex, OpenAI's AI coding agent, which paying users burn through each week. By that evening the post had been viewed around five million times. Picture the view from one of the accounts that saw it. Your Codex dashboard says 85% of this week's allowance is left, with days to go before it refreshes. A reset is hours away, and you do not know what kind. A small, familiar voice starts doing arithmetic. If this wipes the slate, every prompt I send between now and then is free, and leaving it unused is leaving money on the table. If it does not, I am about to throw away the week.
The dilemma has two branches, and OpenAI's help centre defines both. A global reset "is applied directly to eligible usage limits and does not create a saved reset": everyone returns to full at once, and whatever you had not used is simply gone. If that is what is coming, your 85% is melting, and every hour you do not spend it is value you will never see again. A banked reset "is saved to your account and must be applied by you", and applying a full one "refreshes your 5-hour and weekly Codex usage windows and changes your weekly reset date." If that is what is coming, nothing you hold is at risk. Burning tonight just empties the tank you were going to need on Thursday, and forces you to spend the banked reset early to cover the gap. One branch punishes you for waiting. The other punishes you for rushing. The post does not say which branch you are on.
That is why the hint lands as an itch rather than a gift. Unused quota under a possible wipe feels like waste, and waste feels like a personal failing, so the natural response is to go looking for something to spend it on: the refactor nobody asked for, tests for code about to be retired, a second opinion on a design already approved. The developer and writer Max Woolf described the pressure in July, after OpenAI reset the Codex weekly quota six times in two weeks: "I now have to urgently create new ideas to spin down the quota before it inevitably resets again." Notice the verb. Not do the work, create the ideas. Guessing wrong in the other direction hurts too. That same month a user filed an issue on OpenAI's own Codex repository titled, using Sottiaux's nickname, "I spent my usage waiting for Tibo's 7M reset, but the reset never came".
So break the decision down. It turns on three questions. First, is the work real: already agreed, already queued, something you would do this week regardless? Second, what does it cost you to check it: the prompt someone writes, the diff someone reads, the merge someone approves? Third, what is unused capacity worth to you if the reset turns out to be banked, including not having to spend that banked reset early? The rule that falls out of those three fits in one line: spend one more unit only while its value exceeds attention cost + (chance of a banked reset × value of keeping it). Real queued work almost always clears that bar under either kind of reset, so do it and stop worrying. Invented work mostly does not, and the reason is the very word that made the wipe feel tempting.
The quota might be free. Your evening is not. If reviewing invented work costs at least what that work is worth, burning it never pays, even if a global reset is certain. The quota a wipe would have destroyed was a nominal loss, a percentage on a dashboard; the hours spent reading agent output nobody needed are a real one. Above that floor, the less certain you are of a wipe, the more the invented work has to be worth to justify the rush, and the more each unit you keep is worth if the reset is banked, the more certain you need to be. The input that runs out first is never tokens. It is the attention of the few people qualified to judge what the agents produce. A team that burns quota to avoid wasting a resource it rents has spent the one resource it cannot rent to do it.
At 18:23 UTC the answer came: OpenAI was "loading a banked reset into all accounts of our Plus, Pro and Business users". Anyone who burned overnight guessed wrong this time. Had it been global, the burn would still only have paid for work that was real, which is work they should have been doing anyway. That is the quiet lesson of the rush: it is a backlog test. A team with a ranked queue of work worth pointing AI at loses nothing of value under either kind of reset, and a team without one loses under both. The form of the reset belongs to the vendor, which has said its five-hour limit exists partly to "smoothen the load on our compute", and of the six September resets in the timeline above, four were banked. Nobody wins the reset rush. The only position that cannot lose is the one that never needed to join it.
A team that burns quota to avoid wasting a resource it rents has spent the one resource it cannot rent to do it.
A deeper dive
The mechanism turns on what is actually lost in each case. Quota wiped by a global reset feels like waste because it is visible on a usage dashboard as a percentage that went unspent, but its value to the business was only ever what it could have been converted into, and conversion is the expensive part. Every unit of agent output has to be read, judged and either merged or discarded by a person, and that person's evening is drawn from the same small pool of senior attention that approves releases, answers customers and decides architecture. When the work is invented, the ledger is lopsided: the saving is notional and the cost is paid in full. Worse, the cost does not end at review. Changes merged under deadline pressure to use up an allowance join the codebase, the documentation and the test suite, where they must be maintained by people who never asked for them. A rush that looks like capturing free capacity is often a transfer of future maintenance onto the team, at the moment its judgement is most thinly spread.
The second-order trap is that the unit being hoarded is not stable, and the vendor controls every term of it. OpenAI's help centre says future resets are not guaranteed. On 29 August Sottiaux wrote that a round of fixes would make usage go between 10% and 50% further, depending on how Codex is used, which means a week of quota was worth something different on the thirtieth than on the twenty-eighth. Whether a reset is global or banked, when it lands and which plans receive it are all decided on the other side of the table, by a vendor that has said its five-hour limit exists partly to smooth the load on its compute. A team that plans its sprints around those decisions is optimising a quantity it cannot measure against a schedule it cannot see. The durable asset is the one the vendor cannot change: a ranked view of which work deserves agent time, what that work is worth, and who is qualified to judge it. That is a judgement about where to point AI, and most organisations have never written it down.
Work with CLRT
CLRT builds what the reset rush reveals to be missing: a ranked backlog of work worth pointing AI at, costed in the attention it will take to review, so that spare capacity, whether it arrives as a reset, a new model or a larger plan, always has somewhere valuable to go. We work with leadership to decide which workflows deserve agent time, who judges the output, and what must never be invented just to use an allowance. If you want to know where your own teams should point AI before the next hint arrives, start with CLRT Ascent at ascent.clrtstudio.com, or talk to CLRT directly.
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
