On June 12, OpenAI gave Codex users a reset they could save for later.
It called the feature a banked reset. Instead of waiting for an unexpected quota refresh, a user could hold the reset until a project genuinely needed it. On the same day, OpenAI launched a two-week referral promotion for Plus and Pro users. The invited person had to send a first Codex message before either side received a banked reset.
Banked reset announcement\ Referral announcement
That could have remained a small product benefit.
The following 49 days made it part of a much larger pattern.
Tibo Sottiaux, whose X bio identifies his work as Codex and ChatGPT at OpenAI, began appearing frequently in public conversations about quota, resets, and usage rules. In late June, Codex users reported unusually fast consumption. The team reset usage several times while investigating possible sources such as auto-review, subagents, and background suggestions. It later issued another banked reset after the fix.
On July 10, ChatGPT Work launched for long-running, multi-step tasks. It covers research, documents, spreadsheets, presentations, websites, and files. Codex continued to cover code, repositories, terminals, and technical work.
Two days later, OpenAI temporarily removed the five-hour limit for Plus, Business, and Pro users. More resets followed. So did banked resets, active-user milestones, and a public request for feedback on the limit-free experience. On July 29, Sottiaux announced another reset and said that the five-hour limit would return the following day.
Temporary five-hour removal\ User feedback request\ Five-hour return notice
Quota rules had become public events.
Users watched Sottiaux's posts. Communities reminded one another when a reset might arrive. Some people began planning around a possible reset window: should a task run today, or should the remaining quota be held for a larger job?
The platform distributed quota. Users reorganised work.
Why did 6M become 9M in four days?
On July 13, Sottiaux said Codex had reached 6M active users. The next three days brought 7M, 8M, and 9M. Each million-user milestone was accompanied by a reset or a banked reset.
The sequence is striking. It also raises a basic question: what changed during those days?
Several launches overlapped. On July 9, OpenAI released GPT-5.6 across ChatGPT, Codex, and the API. ChatGPT Work began a phased rollout to paid plans. On July 16, desktop updates brought the Chat, Work, Codex, and Projects experience to all plans.
GPT-5.6 release\ ChatGPT Release Notes
At the same time, the platform was issuing resets, removing the five-hour limit, and making banked resets available on web and mobile. A model release, a broader knowledge-work agent, new entry points, wider access, and dense usage benefits landed in the same week.
Those conditions can account for a strong pulse of activity. They cannot account for the source of every reported million users.
The public posts do not define the active-user metric or break out Codex from ChatGPT Work. A newly launched Work product, phased access, returning users, and reset-driven usage could all affect the count. The honest conclusion is narrower: the milestones sit inside an unusually dense release and access window. They should not be drawn as a smooth curve of organic new-user growth.
- 01 · Jun 12
Banked resets and referrals
Quota could be saved, shared, and activated by a first real task.
- 02 · Jun 27–30
Unexpected usage and repeated resets
An incident put supply rules, remedies, and public explanations in view.
- 03 · Jul 9–10
GPT-5.6 and Work launch
Model access, entry points, and task scope expanded in the same week.
- 04 · Jul 13–16
No 5h cap, resets, and 6M–9M
The milestones sit inside an overlap of releases, access, and benefit changes.
- 05 · Jul 18–21
Paid-user reset and competitor actions
Both companies adjusted the working capacity users could access in the same window.
- 06 · Jul 28–30
Work reset and the 5h return
A short-cycle boundary reappeared after a period of intense use.
A second set of supply actions appeared at Anthropic
This was not an attempt to identify permanent strategies for OpenAI and Anthropic. It is a comparison of public actions within the same window.
On July 7, Anthropic began bringing Claude Cowork to the web and mobile, starting with a Max beta. Work could continue across devices and in the cloud after the user closed a laptop. Anthropic also extended Cowork's double usage limit through August 5.
The timing put Cowork next to ChatGPT Work. Both companies were giving agents a wider role in file-based and knowledge work, with more ways to continue a task across devices and sessions.
On July 12, Anthropic extended promotional Fable 5 access on paid plans and a 50 percent weekly Claude Code allowance increase. On July 18, it announced a clearer tiering model: Max and Team Premium received Fable 5 in their included allowance, while Pro and Team Standard continued through credits and received a one-time $100 credit.
July 12 official post\ July 18 official post
The weekly Claude Code increase was extended through August 19. The rolling five-hour limit remained in place.
For that period, Anthropic gave users a clearer calendar of additional weekly capacity, model access, plan tiers, and a retained short-cycle boundary. OpenAI's public actions changed capacity more frequently through resets, banked resets, temporary limit removal, and Work's arrival in the same agentic usage structure.
These are observations within a short window. The longer question sits elsewhere: which platform gives a user enough working capacity, and enough confidence in next week's capacity, to receive more of that user's real work?
A real agent user is expensive
Traditional product growth metrics concentrate on acquisition, activity, and retention.
Agent products add another cost curve. A person who looks at Codex a few times is inexpensive. A person who starts working with it may ask it to read a repository, edit dozens of files, run tests, inspect failures, and try again. A Work user may upload a set of files, ask for research or a report, revise the output, and continue the task.
The more successfully a user hands work to an agent, the more inference, context, tool calls, and peak capacity the platform must provide.
That changes what growth teams need to care about. Account creation is a weak signal. Opening the product is also a weak signal. A first real task matters more. So does a user returning next week with a larger task, or trusting the product with an important project.
Resets, credits, plan tiers, five-hour limits, weekly allowances, and banked resets all help answer the same operational question: how much working capacity is released to whom, when will they use it, and what happens when they reach the boundary?
A heavy user can be both a valuable subscriber and a difficult account to cover with a fixed monthly price.
A reset first gives someone a chance to continue
It would be too easy to treat every reset as a carefully engineered growth experiment.
The late-June resets had an explicit incident context. Users were consuming quota unexpectedly. The team investigated, fixed the problem, and returned working capacity. In that situation, the reset was first a remedy.
Resets also make sense around a model launch. If a user reaches a cap after two quick attempts, the product has little chance to demonstrate what it can do on a difficult repository or a task that needs several iterations. The user needs enough runway to retry, revise, and move a stalled project.
The user experience changes here.
I have called it quota regret. Nothing has been taken away. Yet once a reset is announced, the work that was not handed to an agent yesterday can suddenly feel like a missed option. A stubborn bug might have been worth another attempt. A neglected project might have been worth reopening. Some work completed manually might have been worth testing with Codex first.
Banked resets make that feeling more deliberate. They turn a temporary benefit into a reserve. Users begin deciding whether to save it for a late-month project, a repository that needs repeated tests, or an urgent job that has not appeared yet.
Quota becomes something people manage as working capacity.
A referral promotion asks for the first handoff
The June referral promotion had a precise activation condition.
Accepting an invite was not enough. The invited person had to send a first Codex message before the reward took effect.
That small action sits much closer to an agent product's real goal than registration does. A person can install an app and browse. Someone who hands Codex even a small task has completed a first work handoff. Whether that person later hands over more tasks is where retention, upgrades, and the contest for work allocation begin.
The limited reset reward later ended. Referral access and eligibility have varied by account, region, and plan. Current official terms still allow referral promotions, while specifying that rewards, quantities, expiration dates, and eligibility can vary by campaign. The terms also prohibit self-referrals and attempts to bypass the conditions through aliases or linked devices.
Codex referral promotion terms
Chinese communities also discussed paid invitations, assisted registrations, and trading reset vouchers. This is an observation about community behaviour. It cannot establish why OpenAI changed any particular promotion.
It does show a practical issue. A benefit that can be converted into extra working time can quickly acquire a market price. The platform's cost includes the possible compute behind that time. An account acquired for a reward and an account that repeatedly hands work to an agent have very different value and cost profiles.
Work broadens the range of work OpenAI can receive
Codex is easily understood as a developer product. Work widens that frame.
OpenAI describes Work as a tool for researching topics, analysing information, writing documents, building spreadsheets, creating presentations, producing reports, websites, and file-based deliverables.
Official ChatGPT Work and Codex guide
The usage structure matters as much as the task list. OpenAI's help page says that Work follows Codex's usage structure. Voice-initiated Work and Codex tasks also draw from the same agentic usage and credit pool.
Quota is no longer only a developer rule. A person preparing a research report, organising client material, analysing a spreadsheet, or writing a proposal now encounters the same questions about remaining capacity, long-running tasks, credits, resets, and plan limits.
Codex and Work receive different kinds of tasks. Both compete for the same decision: when the next piece of work arrives, will the user hand it over?
OpenAI
Frequent adjustments in this window
Releases, resets, banked resets, temporary limit changesAnthropic
Visible rule changes in the same window
Weekly allowance boosts, five-hour boundary, plan tiersUsers reorder the work in their hands
Which tasks go to an agent today, which wait for the next cycle, which go to a cheaper model, and which remain human work.The five-hour limit was also a work boundary
When I saw the July 29 notice that the five-hour limit would return, my reaction was relief.
That may sound odd. Users had been asking for more quota and fewer restrictions. During the limit-free period, more projects could move continuously.
Over time, I began to feel that the work could consume the week.
With a five-hour boundary, I plan. If a complex project is coming later in the week, I spend less quota on small tasks. I decide which task needs Codex, which can go to a cheaper model, and which I should do myself. Scarcity creates a reason to stop and sort.
When the boundary disappears, a remaining weekly allowance can keep pulling a project forward. Open another repository. Add one more feature. Hand over another document. Try once more.
As agents can handle more, users discover more work that could be delegated. Work can turn into a process of advancing every project for as long as quota remains. The line between a project and a rest period gets harder to protect.
When the five-hour limit returns, users regain a pause point. They can reserve capacity and decide again what belongs with Codex, what can wait, and what should remain human work.
One change remains. Users have experienced how quickly a project can move without a short-cycle boundary. A plan that once felt sufficient may now feel tight. Someone who treated Codex as an occasional tool may begin to consider Pro, credits, or a higher allowance.
We do not know the internal goal behind the temporary limit removal. Public explanations have referred to incident handling, efficiency improvements, user milestones, and celebration. The user-side result is easier to see: a period of unusually open capacity can raise expectations about what a week's work can contain.
01
Five-hour boundary in place
- Save capacity for complex work
- Sort tasks across models and people
- A clear pause point
02
Short-cycle boundary removed
- Projects continue while quota remains
- More tasks move forward early
- Work can spill into rest time
03
Boundary returns
- Planning and pause return
- Users have experienced a faster pace
- The old plan limit feels more visible
Platforms seek work. Users also choose where work goes.
Across these 49 days, OpenAI made Codex and Work easier to encounter. It gave some people a reason to complete a first task. It gave existing users periods of intensive project progress. It adjusted supply, handled an incident, observed user reactions, restored a boundary, and made higher-frequency demand more visible to credits and higher plans.
That is growth. It is more than acquisition.
The asset at stake is a place in the user's working process. When a new project appears, a user can hand it to Codex, Claude Code, ChatGPT Work, a cheaper model, or a human colleague. That choice determines how much real work a platform receives.
Model capability helps a user try an agent. Predictable capacity, clear rules, and the ability to continue a task help a user trust it with important work.
A reset can make room for several tasks today. A stable, understandable supply of working capacity is what earns a longer place in the user's work.