A “weekly allowance” sounds like something you should be able to spread across seven days.
With a high-load AI agent, that is not necessarily how it feels. A few concentrated hours can consume a large part of the weekly pool, while the dashboard still says the next reset is days away. It can feel like buying a weekly theme-park pass, riding for two hours, and then hearing: “That is all for this week. Please come back next week.”
The key point is that Work/Codex allowance, including Astra, is not measured simply by how long the app is open. OpenAI says usage varies with the model, task, input and output size, reasoning setting, Fast mode, and multi-step work.[1] Depending on the plan, both a five-hour window and a weekly window can apply, and you need allowance remaining in both.[1]
There is also an important caveat: not every sudden drop should automatically be treated as normal. In September 2026, several Pro 20x users reported unusually abrupt allowance drops, and at least one later reported that the displayed allowance had been corrected upward.[2] Heavy real usage and a possible accounting/display problem are different hypotheses.
1. “I only used it for two hours” does not mean the compute load was small
OpenAI explicitly says the same task can consume different amounts depending on the model, and that larger inputs and outputs, higher reasoning, Fast mode, and multi-step tasks can increase usage.[1]
So two hours of human wall-clock time can still contain a huge amount of machine work: reading a large repository, searching, editing many files, running tests, reading failure logs, and iterating again.
Current Codex Pricing estimates local messages per five-hour period at 25–225 for Astra, 50–500 for Sol, 125–1,000 for Terra, and 1,250–10,000 for Luna on Pro 5x. On Pro 20x, the ranges are 100–900, 200–2,000, 500–4,000, and 5,000–40,000 respectively.[3] These are estimates, not fixed message caps, and weekly limits can also apply.[1][3]
That makes model choice a much larger lever than simply shaving a little reasoning effort off every task.
Astra is smart. Astra is also not known for economy-car fuel consumption.
2. A five-hour window plus a weekly pool creates the “weekly pass, instant closing time” feeling
Work and Codex can have both a five-hour allowance window and a weekly allowance. A new five-hour window starts with the first request after the previous window ends, while the weekly pool controls total included work over the week.[1]
Crucially, OpenAI says you may hit the five-hour limit before five hours have actually elapsed.[1]
The same intuition problem appears with the weekly pool. The “100%” gauge looks like seven days of fuel, but it is closer to a quantity of compute you may spend during that weekly period. If you spend aggressively on Monday, Tuesday through Sunday can become a waiting room.
From a user-experience perspective, a daily refill or rolling recovery would feel more natural for bursty workloads. That is a design preference, not the current documented policy.
3. The Pro 20x signup pause is real, but it is a pause, not a shutdown
On September 10, 2026, OpenAI temporarily paused new signups and upgrades to the $200 ChatGPT Pro 20x plan.[4]
Existing Pro 20x subscriptions are not affected and can continue renewing. The $100 Pro tier remains available.[4] So the accurate description is a temporary pause on new Pro 20x subscriptions and upgrades, not the end of the plan.
The official FAQ confirms the pause but does not itself give a detailed technical reason. TechCrunch, however, reported comments from OpenAI product leader Thibault Sottiaux saying that the highest-usage Pro tier put the most strain on the systems, and described Astra demand as unprecedented.[5]
That distinction matters:
- Officially confirmed: new Pro 20x signups/upgrades are temporarily paused.[4]
- Reported explanation: unusually high Astra demand and system-capacity pressure, based on an OpenAI executive’s public comments.[5]
The GPUs are not literally on fire. Probably. But stopping new users from entering the highest-usage tier is a very visible reminder that compute capacity is finite.
4. Do not explain every crazy meter drop with “Astra is expensive”
On September 9–10, users in the OpenAI Developer Community reported abrupt weekly-allowance drops, including one report of 85% of a Pro 20x weekly allowance disappearing in about two and a half hours and another report of Astra Light consuming 20% in a short period.[2]
The interesting part is that the first user later said the allowance had been corrected back to 81% remaining.[2]
That means at least some dramatic readings may involve delayed accounting, display corrections, or another meter issue.
OpenAI Status also recorded an incident with elevated errors affecting some ChatGPT Work users on September 10.[6] There is no confirmation that this incident caused allowance-accounting problems, so the two should not be merged into one claim.
A practical rule is simple: a gradual decline fits a heavy-work explanation better; a sudden teleport-like drop is a reason to preserve screenshots, model/settings information, and usage history and consider an accounting issue too.
A premium model does not make a teleporting meter sacred.
5. Why lowering reasoning may help without transforming the economics
OpenAI says lower reasoning can be a useful starting point for stretching allowance, but a reasoning level does not set a fixed usage amount for a task, and higher reasoning does not always produce a better result.[1]
If the model still has to read a huge repository, the input remains huge. If it still performs many tool calls, the workflow remains multi-step. If it returns large patches and logs, the output remains large.
So using Light can help, but it does not turn Astra into a low-cost model.
Backing off the throttle on a race car does not turn it into a compact commuter car.
The official five-hour estimates also expand dramatically as you move from Astra to Sol, Terra, and Luna.[1][3] For routine work, changing the model can matter more than fine-tuning the reasoning setting.
6. If the goal is a deliverable, measure deliverables
If the goal is “use the premium model for as many hours as possible,” a short weekly allowance is obviously frustrating.
If the real goal is finished code, research, writing, or repairs, the useful metrics are different:
- first-pass success rate,
- bugs and missed requirements left behind,
- test pass rate,
- number of rework loops,
- human correction time,
- total time to a finished deliverable.
If Astra and Sol produce nearly the same final artifact for a task, routine Astra use becomes hard to justify. If Astra identifies a root cause in one pass while another model needs three retries, a short Astra run can still be the cheaper option in outcome terms.
Measure compute per finished artifact, not prestige-model hours.
7. Astra works better as the difficult-problem specialist than the employee who does everything
OpenAI itself positions Astra for difficult bugs, unfamiliar problems, complex research and analysis; Sol for feature implementation and synthesis; Terra for everyday work; and Luna for extraction, categorization, and short repetitive edits.[1]
An output-first split therefore looks reasonable:
- Luna / Terra: extraction, formatting, routine edits, simple file changes, repetitive processing.
- Sol: normal implementation, research synthesis, moderately difficult debugging, longer professional tasks.
- Astra: unclear root causes, architecture decisions, cross-system problems, and final decisions where a mistake is expensive.
Do not keep the ace pitcher on the mound from the first inning every day and then act surprised when the bullpen is empty by the weekend.
8. The deeper problem is not only allowance size; it is the mismatch between burn speed and reset cadence
A model capable of spending a lot of compute in a short burst is naturally awkward when paired with a weekly reset.
The more someone likes to finish heavy work in one concentrated session, the easier it is to use the week’s pool early and then discover a new hard problem later with no premium capacity left.
So the frustration is not just “the quota is too small.” It is also the asymmetry of being able to consume a weekly entitlement in hours while recovery is measured in days.
OpenAI currently documents saved and purchased resets for some eligible accounts, but it does not describe a permanent switch to daily replenishment.[1]
If a weekly pass can be spent in two hours, wanting the gauge to recover a little every day is understandable.
For now, the practical answer is straightforward: use light models for light work, bring in Astra for the hard part, and choose the cheaper path whenever the finished output is essentially the same.
Astra is not the whole theme park. It is the roller coaster. Use it where the roller coaster is actually the point.
Sources
- 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
