A weekly limit sounds like something that should last you the whole week.
With heavy AI agent work, it often doesn't. A few hours of serious use can eat most of the weekly allowance, and the usage page still shows days left until the reset. It feels like buying a weekly theme-park pass, riding for 2 hours, and being told: "That's it for this week. See you next week."
Short version: your Codex and Work allowance (Astra included) doesn't depend on how long the app is open. OpenAI says it changes with the model, the task, how big the input and output are, the reasoning setting, Fast mode, and multi-step jobs.[1] On many plans there is also a 5-hour limit on top of the weekly one, and you need room left in both.[1]
One catch: don't write off every sudden drop as "that's just how it works." In September 2026, several Pro 20x users reported their allowance falling off a cliff, and at least one saw the number go back up afterward.[2] So "the work really was heavy" and "the meter may have miscounted" are two different possibilities.
1. "I only used it for 2 hours" doesn't mean the job was small
OpenAI says the same task can use different amounts depending on the model. Bigger inputs and outputs, higher reasoning, Fast mode and multi-step work all push usage up.[1]
Two hours on your clock can hide a lot of work on the machine's side: reading a huge repository, searching it, changing many files, running tests, reading the error logs, and trying again.
Codex Pricing currently estimates local messages per 5 hours on Pro 5x like this: Astra 25–225, Sol 50–500, Terra 125–1,000, Luna 1,250–10,000. On Pro 20x it's 100–900, 200–2,000, 500–4,000 and 5,000–40,000 in the same order.[3] These are rough ranges, not a guaranteed message count, and the weekly limit applies as well.[1][3]
So the biggest lever is not a slightly lower reasoning setting. It's which model you hand the job to.
Astra is brilliant. It also drinks fuel like a luxury car.
2. A 5-hour limit plus a weekly limit gives you "weekly pass, park closes right away"
Work and Codex can have a 5-hour limit and a weekly limit at once. A new 5-hour window opens with your first request after the last one ends. The weekly limit covers everything you use across the week.[1]
Don't read "5 hours" as a promise you can work for 5 hours. OpenAI itself says you can reach the 5-hour limit before 5 hours have passed.[1]
The weekly meter fools people in the same way. "100%" looks like fuel for 7 days, but it's closer to a total amount of computing you can spend within the week. Sprint on Monday and you may spend Tuesday to Sunday waiting.
Many users would find a small daily refill, or a rolling recovery, more natural. That's my opinion on the design, not how it works today.
3. The Pro 20x pause is real, but the plan isn't cancelled
On September 10, 2026, OpenAI paused new signups for the $200-a-month ChatGPT Pro 20x plan, and also upgrades to it from lower plans.[4]
If you already have Pro 20x, nothing changes and you can keep renewing. The $100 Pro plan is still open to new customers.[4] So the right way to put it: a temporary pause on new Pro 20x signups, not the end of the plan.
The official FAQ doesn't spell out why. TechCrunch quoted Thibault Sottiaux, OpenAI's product lead, saying the Pro 20x users put the most load on the systems, and it described demand for Astra as unprecedented.[5]
So keep two things apart:
- Confirmed by OpenAI: new Pro 20x signups and upgrades have been paused since September 10, 2026.[4]
- Reported reason: heavy Astra demand and pressure on capacity, based on what an OpenAI executive said publicly.[5]
The GPUs aren't literally on fire. Probably. Still, closing the door on the top-usage plan shows pretty clearly that computing power isn't unlimited.
4. Not every sudden drop is just "Astra is expensive"
On September 9–10, people on the OpenAI Developer Community reported sharp drops: one said 85% of a Pro 20x weekly allowance vanished in about 2.5 hours, another said Astra Light used 20% in a short time.[2]
The interesting part: the first user later added that the allowance had gone back up to 81% remaining.[2]
So at least some of these cases may come from delayed counting, a display fix or another meter problem.
OpenAI Status also logged errors affecting some ChatGPT Work users on September 10.[6] Nobody has confirmed that this caused the counting problems, so don't treat them as one story.
A simple rule of thumb: a slow, steady drop probably means heavy work. A sudden jump probably deserves a closer look — save screenshots, the model, the settings, whether Fast mode was on, and your usage history, and keep a counting error on the list of suspects.
"It's a top-tier model" doesn't excuse a meter that teleports.
5. Why switching to Light doesn't always help much
OpenAI says a lower reasoning setting is a good first step to make your allowance last. But the setting doesn't fix how much a task costs, and turning it up doesn't always give better results.[1]
Here's why that matters. Lower the reasoning and a huge repository is still a huge input. Dozens of tool calls are still a multi-step job. Big diffs and long logs are still a big output.
So Light is not pointless, but it doesn't turn Astra into a cheap model.
Easing off the gas in an F1 car doesn't give you a compact car's mileage.
The official 5-hour estimates also grow a lot as you go down from Astra to Sol, Terra and Luna.[1][3] For routine work, choosing a different model can do more than tweaking the reasoning setting.
6. If you want a finished result, track results, not hours
If your goal is to run a premium model for as many hours as possible, a short weekly limit will always feel unfair.
But if what you really want is finished code, an article, research or a fixed bug, then different numbers matter:
- how often it got it right on the first try
- how many bugs or missed requirements were left
- the test pass rate
- how many rounds of redoing it took
- how long you spent fixing things by hand
- total time until the final result
If Astra and Sol give nearly the same final result, there's little reason to use Astra every day. But if Astra finds the root cause in one go while another model needs 3 tries, even a short Astra session is worth it.
Count the computing cost per finished result, not the cost per hour.
That's the most practical way to look at it.
7. Use Astra as the specialist for hard problems, not the worker who does everything
OpenAI's own descriptions put Astra on hard bugs, unfamiliar problems and complex analysis. Sol is for building features and pulling research together, Terra for everyday changes, and Luna for repetitive jobs like extracting, sorting and short edits.[1]
If you plan around results, this split is easy to follow:
- Luna / Terra: searching, extracting, formatting, routine fixes, simple file edits, repetitive work.
- Sol: ordinary implementation, pulling research together, moderately hard debugging, longer work tasks.
- Astra: bugs with no obvious cause, design decisions, problems that span several systems, and final calls where a mistake is costly.
Don't run Astra on top settings all the time. Bring in the ace only when you're stuck.
If you send your closer out in the first inning every game, of course he's gone by the weekend.
8. The real issue isn't just a small limit; it's burning fast against a slow reset
A model that can burn through a lot of computing in a short burst doesn't fit well with a weekly reset.
The more you like to knock out heavy work in one go, the easier it is to empty the week's allowance early. Then a new hard problem shows up and you're stuck until the reset.
So the frustration isn't only about the total being small. It's the mismatch: you can spend a week's worth in a few hours, but getting it back takes days.
OpenAI says some accounts may get saved resets or purchased resets, but it hasn't said it is moving to a steady daily refill.[1]
If a weekly pass can be used up in 2 hours, it's natural to wish the meter would fill back up a little every day.
Even so, if you care about results, the practical answer today is simple: give light jobs to light models, bring in Astra for the hard part, and pick the cheaper model whenever the result is the same.
Astra isn't the whole theme park. It's the roller coaster. It's great where you actually need it, but you don't have to ride only the roller coaster from opening to closing.
References (6)
- OpenAI Help Center — Managing usage with GPT-6 Astra in Work and Codex. Retrieved 2026-09-14 help.openai.com
- OpenAI Developer Community — “Insane usage burn rate swing on Pro 20x, potential bug.” User reports from 2026-09-09 to 2026-09-10; anecdotal evidence only. Includes a report of a later allowance correction community.openai.com
- OpenAI / ChatGPT — Codex Pricing. Retrieved 2026-09-14 chatgpt.com
- OpenAI Help Center — About ChatGPT Pro tiers. Retrieved 2026-09-14. Confirms the temporary pause on new Pro $200 / Pro 20x signups and upgrades beginning 2026-09-10, while existing subscriptions are unaffected help.openai.com
- TechCrunch — “OpenAI puts Pro subscriptions on hold due to Astra demand.” Published 2026-09-10. Reports public comments from OpenAI product leader Thibault Sottiaux about system strain and unprecedented Astra demand techcrunch.com
- OpenAI Status — Elevated errors affecting ChatGPT Work. Incident on 2026-09-10; resolved. This source does not establish that the incident caused usage-accounting issues status.openai.com
