You want to move to the higher plan and run more AI work. The upgrade gets stuck. Meanwhile, usage resets are being handed out around you.
At some point the joke writes itself:
“Just lend me Tibo’s computer. Unlimited, please.”
No, this is not a proposal to borrow somebody else’s account or internal permissions. It is what happens when GPT-6 Astra turns Work and Codex allowance into something that feels less like a message quota and more like research capacity.
The interesting question is not merely “Is 20x better than 5x?”
The real multiplier is allowance × resets × uninterrupted human time.
1. Five-second answer: 20x is four times 5x, and a reset can restore a much larger absolute pool
OpenAI currently describes two ChatGPT Pro tiers:
- Pro $100: 5x the Plus usage allowance
- Pro $200: 20x the Plus usage allowance
So Pro 20x provides four times the allowance of Pro 5x, not twenty times as much as 5x.
OpenAI’s Astra guidance shows the same ratio in its estimated local messages per five-hour period: 25–225 for Pro 5x and 100–900 for Pro 20x. These are estimates, not fixed message limits. Task size, input and output length, reasoning effort, Fast mode and other factors change consumption.
A Full reset can refresh the eligible five-hour and weekly Work/Codex windows.
That means one reset on a 20x plan can restore a larger absolute amount of included work than one reset on a 5x plan.
Same reset button. Bigger tank.
2. What does the “20” mean? Both numbers are relative to Plus
Use a simple unit model. Call the Plus allowance 1.
- Plus: 1
- Pro 5x: 5
- Pro 20x: 20
Moving from 5x to 20x adds 15 baseline units and makes the total four times larger.
That distinction matters because the name “20x” sounds enormous in isolation. For a 5x subscriber, however, the upgrade is a 4x increase.
Whether that is valuable depends on actual demand.
If someone already drains 5x repeatedly, quadrupling a scarce resource can be dramatic. If somebody uses only half of 5x, upgrading mostly creates a larger pile of unused allowance.
Unused compute has impressive theoretical capacity and exactly zero practical output.
3. Why does a Full reset look so much “fatter” on 20x?
OpenAI says a Full banked reset refreshes eligible five-hour and weekly Work/Codex usage windows and changes the weekly reset date. It is a restoration mechanism, not a permanent increase in the plan.
Now take a deliberately simplified thought experiment.
Suppose a Pro 20x user fully consumes the relevant allowance, then can successfully use two independent Full resets after meaningful depletion.
normal 20
+ first restored 20
+ second restored 20
= 60
Measured in our fictional Plus-baseline units, that looks like 60 units of capacity during the observation window.
Compare that only with one untouched Pro 5x period and you get 60 ÷ 5 = 12x.
That is where the tempting headline comes from: “20x plus resets is basically 60x!”
But it is not an apples-to-apples plan comparison.
If the 5x user receives the same two resets and also fully uses them:
5 × 3 rounds = 15
20 × 3 rounds = 60
60 ÷ 15 = 4
The plan ratio remains 4x.
Also, a reset can change the weekly reset date, so treating a fixed Monday-to-Sunday calendar week as an accounting box can be misleading.
The honest conclusion is narrower: a Full reset is especially valuable in absolute terms to a power user who can actually consume the restored 20x pool.
4. September 2026 really did feel like a week run by “The Reset Company”
The Astra launch period was unusually reset-heavy.
OpenAI says eligible existing Plus, Pro and Business subscribers in good standing received a banked reset on September 3 and another on September 4, 2026. On September 7, eligible users received a separate automatic global reset.
On September 9, OpenAI Status reported an incident in which some Codex users experienced unexpected usage-limit resets. The incident was mitigated and marked resolved that day.
Afterward, Thibault “Tibo” Sottiaux, who leads Codex at OpenAI, publicly said that people who used a banked reset during the affected window, when it did not fully apply in Work and Codex, would receive another reset and an apology email.
From the user side, the scene starts to look like this:
Tibo has returned to the monetary printer, except the currency is compute.
Tibo Central Bank is open.
There is one crucial caveat. OpenAI explicitly describes banked resets as promotional benefits and says future resets are not guaranteed.
So “I expect roughly two free resets every week forever” is not a sound subscription model.
A launch festival is a festival. It is not a pension system.
5. Reset value depends on how empty your tank was before the reset
The useful value of a reset is not simply the number of reset events. It depends on how much allowance had been consumed beforehand.
If 90% remains and the account returns to 100%, only about 10 percentage points of additional room were recovered.
If the pool is near empty, a successful full restoration can be far more valuable.
So the same promotional event looks very different for different users:
- a light 5x user may barely notice it;
- a heavy 5x user can gain a substantial extra run;
- a heavy 20x user can recover a very large absolute amount of capacity.
This gives a better upgrade test than hype.
If you already exhaust the lower tier, your demand has been observed rather than imagined.
If the lower tier remains half unused, “I may need it someday” is much closer to FOMO.
6. Once compute expands, the bottleneck can move from the AI to the human’s free time
Quadrupling AI allowance does not quadruple results automatically. Somebody still has to choose worthwhile jobs, inspect outcomes and redirect the system.
Ask Astra to explore.
Review the output.
Reject a weak hypothesis.
Add missing data.
Repair the environment.
Turn a successful procedure into a script.
Those activities need cognitive room, not just spare tokens.
This creates a strange AI-era inversion:
Access to a powerful model can become less scarce than uninterrupted time to keep feeding it useful work.
An extra hour or two of overtime does not only remove those clock hours. It can also consume recovery time and make it harder to start work that requires deep context, experimentation and repeated iteration.
Thirty free minutes is plenty of time to answer email.
It is a terrible timeslot for building a research lab.
7. “Slack time is powerful” is more than a celebration of doing nothing
There is empirical work directly concerned with slack time and innovation.
Agrawal, Catalini and Goldfarb studied the timing of innovative projects on a major crowdfunding platform and documented a sharp increase when colleges were on break. Their evidence is consistent with slack time helping not only ideation, but also mundane execution work such as administration, planning and promotion that turns an idea into something actually launched.
That mechanism maps surprisingly well onto serious AI use.
The glamorous part is “generate a new idea.” The time-consuming part is often:
- setting up the environment;
- cleaning data;
- writing tests;
- preserving logs;
- documenting commands;
- making runs reproducible;
- fixing boring failure cases.
Pencavel’s work on historical munitions production also found a nonlinear relationship between hours and output: beyond a threshold, additional hours produced diminishing increments in output.
A related analysis examined recovery between workweeks and reported results consistent with the idea that long workweeks create fatigue and insufficient recuperation that can hurt subsequent performance.
These studies are not experiments on modern AI power users, so they do not justify a precise claim such as “one hour of overtime reduces creativity by X%.”
They do support a more modest point: time, recovery and execution capacity matter, and output does not necessarily scale linearly with hours worked.
8. If you buy one month of 20x, buy a research factory rather than a pile of answers
A short burst of expensive compute is wasted if it leaves behind only hundreds of disconnected answers.
The stronger strategy is to convert temporary allowance into infrastructure that can be reused after the month ends.
OpenAI positions Work for longer, multi-step work and finished deliverables, while Codex remains focused on software development. On desktop, Codex can work with local folders, repositories, terminals and developer tools; where allowed, Work can also use local files and desktop apps with permission.
A one-month sprint can therefore target a repeatable loop:
explore
↓
form a hypothesis
↓
implement
↓
run an experiment
↓
evaluate
↓
record failures
↓
generate the next hypothesis
“AI researched 100 candidates for me” is useful.
“AI built a system where one candidate triggers 100 tests and returns a ranked report plus failure logs” is an asset.
The second result keeps producing value when the promotional resets disappear and the subscription returns to a lower tier.
9. Environment construction is how temporary free time becomes persistent capital
A research factory needs more than code that happened to run once.
Useful deliverables include:
- Git history;
- locked dependencies and versions;
- scripted setup;
- a README with exact run commands;
- tests;
- saved experiment conditions and results;
- failure logs;
- stable input and output schemas;
- reproducibility on a replacement computer.
Then a future busy period starts from “review yesterday’s automated results,” not “rebuild the environment from scratch.”
That is the compounding effect.
Permissions still deserve a separate design. Production credentials, secret keys, publishing rights, transfers, live trading and destructive actions should be isolated from exploratory environments. Give the agent the minimum access it needs and preserve human gates around high-impact irreversible actions.
Let AI build the laboratory.
You do not need to leave the vault key on the laboratory desk.
10. Evaluate the extra $100 by hours recovered, not by the word “20x”
OpenAI lists Pro 5x at $100 and Pro 20x at $200, a $100 monthly difference before any local tax, currency or billing variation.
Instead of asking only whether “$100 for AI” sounds expensive, ask how much human time a concentrated month can recover.
Simple cost allocation looks like this:
- 10 hours saved: $10 per hour
- 20 hours saved: $5 per hour
- 40 hours saved: $2.50 per hour
That is not a promise of profit. It is simple division.
The case becomes stronger when the month produces automation that continues saving time later.
A sensible decision rule is:
Upgrade-leaning
- you already hit 5x limits often;
- you can actually consume banked resets;
- you have a defined research or development backlog;
- you currently have a rare block of free time;
- you can leave behind reusable automation and environments.
Skip-leaning
- 5x remains unused;
- “Astra is new” is the main reason;
- there is no defined backlog;
- you lack time to review outputs;
- the economics require free resets to continue on a predictable schedule.
Treat surprise resets as upside.
Do not book them as recurring revenue.
11. Even an upgrade delay has an opportunity cost when timing is the whole point
The value of AI capacity depends on when it becomes available.
Imagine somebody has a long break between jobs or projects and wants to use that temporary window to build a research environment. If an upgrade is delayed for several days by billing or plan-switching problems, the loss is not merely the subscription price difference.
The person loses the overlap between maximum human slack and maximum AI capacity.
A month of 20x later, when normal work has resumed, may not be worth the same amount.
Price pages list dollars and allowances. They do not list the market value of your uninterrupted Tuesday afternoon.
But the real return on the subscription depends on it.
12. Conclusion: in the AI era, uninterrupted slack may be scarcer than tokens
Pro 20x provides four times the base allowance of Pro 5x.
During an unusual launch period with additional resets, power users who can drain their pool can obtain much larger absolute value from each restoration. Future free resets, however, are not guaranteed.
So “buy 20x because Tibo will surely reset it again” is weak reasoning.
The stronger case is this:
You have a temporary block of free time, and you can use Astra, Work and Codex to convert that time into exploration systems, reproducible environments, tests, logs and automation.
Slack that is merely consumed disappears.
Slack used to build a system that keeps working while you are busy can return again and again.
The ultimate AI R&D advantage may not be an unlimited plan.
It may be uninterrupted time used to build a laboratory that no longer requires uninterrupted time every day.
And if you still burn through everything?
Tibo, about that computer. Unlimited, please.
Sources
- OpenAI Help Center, “About ChatGPT Pro tiers” — https://help.openai.com/en/articles/9793128
- OpenAI Help Center, “Managing usage with GPT-6 Astra in Work and Codex” — https://help.openai.com/en/articles/20001516
- OpenAI Help Center, “How banked Codex resets work” — https://help.openai.com/en/articles/20001498
- OpenAI Status, “Investigating unexpected usage limit resets” — https://status.openai.com/incidents/01M23KG62KKK448RN434CQ64Z8
- Tibo Sottiaux, public X post on replacement resets for the affected window — https://x.com/thsottiaux/status/2097752790177370535
- OpenAI Help Center, “ChatGPT Work and Codex” — https://help.openai.com/en/articles/20001275
- OpenAI Forum, “Codex is for Everyone: Why Codex Matters Beyond Code” — https://forum.openai.com/public/events/codex-is-for-everyone-why-codex-matters-beyond-code-fa40puy7wi
- Agrawal, Catalini & Goldfarb, “Slack Time and Innovation,” NBER Working Paper 21134 — https://www.nber.org/papers/w21134
- John Pencavel, “The Productivity Of Working Hours,” Stanford SIEPR — https://siepr.stanford.edu/publications/working-paper/productivity-working-hours
- John Pencavel, “Recovery from work and the productivity of working hours,” Stanford SIEPR — https://siepr.stanford.edu/publications/working-paper/recovery-work-and-productivity-working-hours
