Skip to content
Fully Doxxed

The Token Plateau: What 147 Codex Power Users Say About OpenAI's Next $10 Billion

— AI, OpenAI, Codex, Agents, Data — 13 min read

OpenAI just reported $10 billion in revenue for a single quarter, up nearly 50% from the quarter before. That is a big number, and it's the kind of number that makes people draw a straight line on a log chart and extend it to the moon.

But revenue is an output. The thing underneath it is usage: people (and, increasingly, their agents) burning tokens. And usage is hard to see from the outside. OpenAI doesn't publish it per user, and everything we normally get is an anecdote. Somebody posts a screenshot of a trillion-token month, somebody else complains they hit their limit by Tuesday, and we all update on vibes.

So Claude and I went and got some data. It turns out that some Codex users have made their usage public, one day at a time, for a whole year. We found 147 of those profiles. Here is the headline chart:

Rapid growth… then stagnation
Tokens on an active day for a set of power users. Weekly median across 147 public Codex profiles.
Scale
Dashed: the growth rate fitted to November 2025 through July 2026, carried forward. An illustration, not a forecast. Red: the median week since the July peak. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

Basically, there are two stories in that picture. The first is that usage went completely vertical. The second is that it then stopped. If you flip the chart to a log scale, the break in July is hard to miss.

In this post I'll go through where the data comes from, what it shows, what I think is going on, and what it might mean for OpenAI's next $10 billion. It's a self-selected set of very heavy users, and I'll try to be careful about what that does and doesn't let us conclude.

Where the data comes from

In June, OpenAI launched public profile pages for Codex. If you turn yours on, anyone can visit chatgpt.com/u/your-username and see a card with your lifetime tokens, your streak, your favorite plugins, and a little contribution-graph-style calendar of how many tokens you used each day for the past year.

That calendar is the interesting part. It means that every public profile is a year-long daily time series of somebody's AI usage. Nobody else publishes anything like it.

The hard part is finding the profiles, since there's no directory. So we went looking for links that people had already posted in public: the Wayback Machine's index, X posts where people show off their stats, GitHub READMEs, and a few search engines. That turned up 217 usernames, of which 159 were still public and 147 had a daily usage history.

A few ground rules we followed:

  • We only collected profiles whose owners made them public and whose link had been posted or archived somewhere public.
  • We read each profile once, from an ordinary logged-in browser tab, one request at a time and at least a second apart.
  • Everything below uses made-up code names. Meet Glass Otter and Midnight Daemon. They are real people, but those are not their names.

All told, that's 22,526 days of usage and 6.5 trillion tokens. For scale, that is something like fifty million novels' worth of text. Most of it is not prose, of course. It's agents re-reading the same codebase before changing one line.

Now, the big caveat, and I'll repeat it later: these are not typical users. They are people who use Codex enough to want to show it off. Think of this as a study of the frontier, not the average. The top ten profiles account for 39% of all the tokens.

First, the ramp

In this set, a typical active day went from 7 million tokens last November to 115 million this November. That is 16x in twelve months.

Nobody types sixteen times more than they did last year. What changed is that people stopped chatting with the model and started handing it work. Every token on these profiles is agent usage: tasks that run for an hour instead of a minute, and then several agents running at once.

The aggregate curve hides what this looks like for one person, though. Here are individual users, one panel each:

Same shape, over and over: a steep ramp, then a plateau
Weekly tokens from November 2025 to November 2026 for the 41 profiles with this pattern, twelve at a time. Each panel has its own scale.
1–12 of 41
Gilded Kernel
plateau 710M/day
Velvet Socket
plateau 570M/day
Glass Otter
plateau 560M/day
Crimson Lynx
plateau 480M/day
Humming Lambda
plateau 370M/day
Quiet Stoat
plateau 360M/day
Midnight Osprey
plateau 280M/day
Copper Gecko
plateau 250M/day
Midnight Daemon
plateau 210M/day
Static Osprey
plateau 190M/day
Paper Diff
plateau 73M/day
Quiet Compiler
plateau 40M/day
Code names are used to protect privacy. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

Click through a few pages. It's the same shape over and over: months of almost nothing, a steep ramp that takes weeks, and then a plateau. 41 of the 147 profiles fit this pattern cleanly, and plenty of the others are a messier version of it.

Economists will recognize this. It's an adoption S-curve, the thing you see when a country gets washing machines or a company gets email. The difference is the speed. This S-curve happens inside one person, and the steep part typically takes about three weeks (though a few people took months).

And these plateaus are high. Look at how many tokens these people have used in total:

The typical power user has put 14 billion tokens through Codex
Lifetime tokens of 146 public profiles, log scale. Each dot is one profile, smallest to largest.
Compare with
118 of 146 profiles have used more tokens than a 160,000-word novel, by its author's count (about 2.5B).
10M100M1B10B100B1.0TA novelmedian 13.6B
Source: lifetime counters on public Codex profiles, 10–11 Nov 2026. Project figures are their authors' own counts (fullhoffman.com and theateliermethod.org); both were built with Claude Code, and counting methods differ.

The typical profile in this set has put 13.6 billion tokens through Codex. For comparison, one developer publicly reports shipping an iOS, Android and web app on about 1 billion tokens, and one author reports writing a 160,000-word novel on 2.5 billion. (Both of those used Claude Code, and everyone counts tokens a bit differently, so treat it as a rough yardstick.) 131 of the 146 profiles with a lifetime counter have passed a billion.

Then, the plateau

So where does usage level off?

Where 41 power users leveled off: a median of 268M tokens a day
Tokens per day at each user's plateau, log scale. Grey lines: what each plan holds per day, by outside estimates.
Hover or tap a row to meet the user.
10M100M1.0B$20 plan≈20M$100 plan≈157M$200 plan≈461M$500 plan≈1.3Bmedian 268M
Bars show the range of the plateau weeks. Plan sizes are third-party measurements; OpenAI publishes no token limits per plan. Code names are used to protect privacy. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

The median plateau is 268 million tokens a day. At API prices that would be roughly $50 a day, or about $1,500 a month. These users are almost certainly not paying that. They're on subscriptions, and the plateaus cluster between what a $100 plan and a $200 plan can hold, going by outside measurements. (OpenAI doesn't publish token limits for its plans, so the grey lines come from third parties who measured their own accounts.)

At this point you're probably thinking what I was thinking: the plateau is just the plan limit. People ramp up until they hit the ceiling, and then they sit there with their faces pressed against it. That would be a boring explanation, but boring explanations are usually right.

Except this one doesn't hold up.

It's (mostly) not the plan limit

If you lived at a weekly cap, your weeks would all look alike: each one would come in right around your maximum. So take each plateaued user, find their best week, and ask how the rest of their weeks compare.

Power users are not maxing out: the typical week is 40% of their own best week
Each week as a share of the same user's best week. 41 plateaued users, last 12 weeks, best week left out.
Orange: where most weeks would fall if these users lived at a weekly cap. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

The typical week is 40% of the user's own best week. Only 5% of weeks come within 15% of it. And these people are not taking time off: they're idle on just 4% of days. They show up nearly every day and use well under what they've shown they can use.

We can run a version of this for everybody, not just the clean plateau cases. Guess each user's plan from their best week (being generous and assuming that week included a limit reset), then compare a typical week to that plan's allowance.

Against the smallest plan that fits their best week, the typical user runs at 45% of the allowance
Each dot is one user's typical week as a share of a plan's weekly allowance. The plan is a guess, made from their best week.
0%25%50%75%100%125%150%full weekly allowance$2018 users$10047 users$20037 users$50024 users28%34%56%52%
Users over 100% in a typical week got there with limit resets or more than one account. 138 active users; those needing more than one plan are left out. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

The typical user runs at 45% of the allowance. More than half use under half. Only one in ten uses 80% or more.

That last group is real, and I don't want to wave it away. If you're on X you've met them, because they are loud. But in this set they're a tail, and the body of the distribution is people who pay for a lot of headroom and leave most of it on the table.

But limits still matter at the margin

Here's a fun wrinkle. OpenAI has reset everyone's usage limits 59 times this year: as an apology for outages, to celebrate milestones, and at least once because the head of Codex had slept well. The folks at codex-reset.com keep a log, which means we have a string of natural experiments.

The day after OpenAI resets usage limits, users spend 62% more than in the three days before
Tokens per day against each user's own ordinary day (100), averaged over 18 resets at least four days apart.
Blue: the typical profile. Grey: heavy users (300M+ tokens a day). Reset dates: codex-reset.com (CC BY 4.0). Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

In the three days before a reset, users spend 18% less than on an ordinary day. The day after, they spend 32% more. And remember, this is a set where most people are nowhere near their cap.

This is just how people treat a budget. You get careful at the end of the month even when there's money in the account, and you loosen up on payday. A usage meter doesn't have to hit zero to change behavior. It only has to be visible.

A natural experiment in September

If you looked closely at the first chart, you noticed a hole in it. For three weeks in September, tokens per day fell by half and then snapped right back. I assumed it was a data problem. It wasn't.

Three weeks in September when tokens per day halved
Median tokens on an active day, daily, mid-August to mid-October 2026.
  1. 1Sep 3: Astra launches
  2. 2Sep 10: $200 sign-ups paused
  3. 3Sep 22: Sol and Luna: half the credits
  4. 4Sep 29: GPT-6.1 Sol, sign-ups reopen
Event dates from OpenAI's changelog and codex-reset.com. The link to the dip is circumstantial: profiles do not record which model was used. Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

The dip starts within a day or two of September 10 and ends on September 30. Those dates line up with the launch of GPT-6 Astra, OpenAI's new flagship model, which uses up a plan's allowance much faster per token than the models people had been running (about five times faster than Sol, the cheaper model that followed). On September 10 OpenAI paused new sign-ups for the $200 plan. On September 22 it shipped Sol and Luna at half the credits per token, and on September 29 it shipped an updated Sol and reopened sign-ups with re-cut Pro plans. Usage recovered the next day.

What makes me believe this explanation is who dipped. Split users into thirds by how much they were using before:

  • Light users (around 10M tokens a day): no drop at all.
  • Middle users (around 100M a day): down about 5%.
  • Heavy users (around 400M a day): down about a third, and three in four of them fell.

And nobody stopped showing up. The share of days people were active barely moved. They used Codex just as often and got fewer tokens out of it, which is what you'd expect if a pricier model shrank what a plan could hold.

I can't prove this, since profiles don't say which model was running. But I think the lesson is important: a token is not a stable unit. A plan is a budget of compute, and how many tokens that buys depends on the model. So "most users are far below their limit" is true on average and was probably false for heavy users during those three weeks. The limit is a moving target, and for the heavy tail it does bind.

So why do people level off?

If it's mostly not the plan limit, what is it? Here I'm going beyond the data, so take this as a reading and not a result.

My guess is that the ceiling is the human. An agent needs a task, and somebody has to think of that task, describe it, and then check what comes back. In the ramp, a person is discovering how much of their work they can hand off. On the plateau, they've handed off what they know how to hand off, and the scarce input is no longer tokens. It's their own attention.

If that's right, you'd expect the plateau to jump when the tooling changes how much one person can supervise, and to sit still otherwise. The data is at least consistent with that. Growth in this set came in a handful of bursts, and the bursts line up with new models, new agent features like the desktop app and goals, and looser limits. I can't separate those three, which is why this is a guess.

Growth is slowing, and it's the veterans who slowed

Add everyone up and the slowdown is plain:

Total token growth slowed: doubling time stretched from 4 weeks to more than 14
Daily growth rate of the set's total tokens. Blue: fitted trend with its 95% interval. Grey: measured over rolling 12-week windows.
Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

In February the set's total tokens were growing about 2.6% a day, doubling every four weeks. Now it's 0.7% a day, doubling every fourteen and a half weeks. That is still very fast growth by the standards of any normal business! But it is a different regime. (Some of the early number is an artifact of my set filling in, since people join it over time. Most of the slowdown is real.)

And here's who is still growing:

The longer someone has used Codex, the less their usage is still growing
Growth of the typical user from August to November 2026, by when they started. The same three months for everyone.
+30%
Started before February46 users, now 308M a day
+111%
Started February to April38 users, now 128M a day
+158%
Started May to July30 users, now 117M a day
Source: 147 public Codex profiles, daily token counts, Nov 2025 to Nov 2026. Self-selected heavy users, not a sample of all users.

Over the same three months, people who started before February grew 30%. People who started between February and April more than doubled. People who started between May and July grew 158%. Everyone ramps, and then everyone levels off; the veterans just got there first.

What this means for the next $10 billion

Economists like to split growth into the intensive margin (existing customers using more) and the extensive margin (more customers). For the past year, Codex has had an incredible intensive margin. Each power user became sixteen power users' worth of demand without OpenAI having to acquire anybody.

This data says that margin runs out, user by user, within weeks of the ramp starting. Existing users still account for most of the token growth in this set since August, but it's coming from the newer ones, and each cohort flattens in turn.

So I think OpenAI's next 50% has to come from somewhere else. I see three places.

Making more power users. The 41 people in the ramp chart figured out how to hand real work to agents. Most ChatGPT users haven't. Every person who makes that jump goes through the same steep ramp, so the size of the prize is the number of people who haven't started yet. Expect a lot of product work aimed at making agents easier to start with.

Raising the plateau. If the ceiling is human attention, then anything that lets one person direct more work lifts it: longer-running tasks, agents that manage other agents, better ways to review what came back. Each of those would show up in this data as a new burst.

Charging more per token. September showed that when a better model costs more allowance, heavy users don't quit. They just get fewer tokens. That is pricing power, at least over the people at the top of the distribution.

The bear case is the mirror image: the people most excited about agents have already ramped, the next cohort is less eager, and the plateau stays put. I don't think that's the most likely outcome, but this data can't rule it out.

What this data can't tell you

I've been flagging caveats as I go, but they deserve their own section.

  • It's 147 people who chose to make their profile public. They are heavy users and probably proud of it. Nothing here describes the median ChatGPT subscriber.
  • It's one snapshot. People who quit Codex and took their profile down are missing, so the set is tilted toward the ones who stuck around.
  • There's no plan data. I don't know what anyone pays. Plan sizes are outside estimates, and my plan guesses are guesses.
  • Tokens aren't work. By outside measurements, the large majority of Codex tokens are cached input: an agent re-reading files it has already seen. More tokens doesn't have to mean more got done.
  • "Agents drive it" is an interpretation. The profiles count tokens. They don't count agents.

Still, I think the shape is real, because it shows up person after person: a fast ramp and then a flat line that sits below the limit. The aggregate numbers will keep going up for a while, as new people climb the ramp. But each of them is climbing toward a plateau. Whoever figures out how to raise it gets the next $10 billion.

This started as a thread on X, if you'd like to argue with me there:

© 2026 by Fully Doxxed. All rights reserved.