Build1 publisher3 min readPublished
OpenAI's reinstated five-hour cap puts Codex sprints back on a rolling meter
A dev.to round-up of Hacker News and GitHub threads collects users who hit the returned Plus cap mid-sprint, alongside Anthropic weekly quotas that one user says reset as often as every five days. All of it is user reports.
The Engineer · Build desk
What happened
- kylecorpin wrote on Hacker News that the rolling five-hour cap bricked his project mid-sprint and that he was now forced to upgrade.
- venzaspa said the changes mean you cannot plan usage across the week, because an unexpected reset can waste 80 percent of a remaining allowance.
- The account of all this is a dev.to Voices round-up that lifted quotes verbatim from Hacker News and GitHub threads, each linked with handle, platform and date.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint A weekly number that resets at least every five days fails as a sprint budget, because a team paces against one quantity and gets another.
- decision Teams now choose between buying a higher tier for headroom they may not need and designing agent workflows that checkpoint and resume when a window closes.
- exposure According to qsort, the model behind next week's allowance is also uncertain, so prompt suites and evals tuned to one model sit on a substitutable dependency.
A rolling window and a monthly bucket are different objects. Under a rolling cap, what you can spend in the next hour depends on what you spent in the previous five, so a long agent run leaves less headroom for the one after it. kylecorpin, writing on Hacker News, described what that did to one afternoon: "This draconian 5-hour rolling cap completely bricked my project mid-sprint. Now I'm forced to upgrade. What a sham." [6]
The dev.to round-up steel-mans that complaint itself, noting you can read "forced to upgrade" as a heavy user meeting the edge of a cheap plan [14]. It also says the specific grievance, that the cap landed mid-sprint and turned paid work into a dead stop, recurred all week in different wording [15].
What users keep describing is a planning problem. venzaspa put it this way: "I used to think it was pretty great but now it just means you can't plan your usage across the week. It's a slightly entitled complaint but when you 'waste' 80% of your remaining usage because you get a reset you weren't expecting, there's a ping of disappointment there." [7]
On the Anthropic side, Petersipoi said weekly usage "gets reset CONSTANTLY" and that "The longest I've ever seen it go without a reset is maybe 5 days?" [10]. Take that five-day maximum at face value and a seven-day period contains at least 1.4 resets, so the weekly figure an account paces against spans less than a week [16]. Timing was also counter-intuitive in the other direction: mbreese reported being "hit with limits at 4am Eastern time when I'd expect for not a lot to be going on" [8].
What gets metered can move too. qsort said: "They now give out resets you can choose when to use, but with a 'best before' date. With Anthropic you don't even know what model you'll have access to next week." [9] A prompt suite tuned against one model is exposed to substitution on the same subscription.
All of this is user testimony. The dev.to piece is a Voices round-up that lifted quotes verbatim from Hacker News and GitHub threads, each linked with handle, platform and date [3], and it carries no allowance figure in tokens or requests and no vendor statement. The sample is self-selected: people who hit a wall hard enough to post about it.
For a quota to work as a capacity input, three things would have to be published and stable: the allowance, the reset schedule, and the model behind it. The week's record leaves all three open. OpenAI's five-hour cap for Plus on Work and Codex was lifted earlier in the summer and brought back this week [1], and Anthropic's weekly quotas continued resetting on their own rhythm [2].
Two users reached for AOL. mostertoaster asked whether "we look back at how AI usage is charged today and we equate it with how we had minutes on AOL and how absurd it seems looking back?" [12]. olyjohn went further, saying minutes on AOL were at least honest, and that the same thing today would be called unlimited AOL and still carry a limit [13].
What to watch
- Whether OpenAI publishes the allowance and window for the Plus cap on Work and Codex in its own documentation.
- Whether Anthropic documents a fixed reset cadence for weekly quotas, given the ad-hoc resets with expiry dates that qsort described.
- Whether the reinstated five-hour cap holds for a full quarter, having already been removed once earlier in the summer.