OpenAI Ignored a 44.2 Billion-Token Codex User—Then Locked Him Out
While spending hundreds of millions to shape the AI conversation, OpenAI missed one of its most extreme builders, a broken Codex pricing model, and the evidence sitting inside its own platform.
AN OPEN LETTER TO OPENAI
You Bought the Conversation and Missed the Evidence
To OpenAI:
According to my Codex usage dashboard, I processed approximately 6.8 billion tokens in one week.
My displayed lifetime Codex usage is approximately 44.2 billion tokens.
Your response was not an email.
It was not a request to study the workflow.
It was not curiosity from the Codex team, product research, infrastructure, finance, partnerships, communications, or anyone responsible for understanding what people are actually doing with the technology.
Your response was a usage limit telling me that I could not continue working until July 24, 2026.
That is completely fucking absurd.
And it exposes a much larger institutional failure inside the company currently presenting itself as the center of the artificial-intelligence revolution.
OpenAI publicly celebrated a racing game that Codex created using more than seven million tokens. You presented that number as evidence of long-horizon agentic work: one prompt, millions of tokens, multiple roles, repeated iteration, testing, and a playable result. (OpenAI)
My dashboard recorded approximately 6.8 billion tokens in seven days.
That is roughly 971 times the token count of the Codex demonstration you advertised as an achievement—in one week.
I understand that raw token totals are not identical to internal compute cost. Cached input, ordinary input, output, model selection, and speed settings consume credits differently. OpenAI’s own rate card makes that explicit. (OpenAI Help Center)
That qualification does not weaken my point.
It strengthens it.
Whether 6.8 billion tokens represented unusually efficient cached context, unusually expensive generation, or some mixture of both, it was an unmistakable behavioral outlier.
It demanded one question:
Who the fuck is doing this, and what are they building?
Apparently, nobody at OpenAI was curious enough to ask.
You Cannot Pretend the Signal Was Impossible to Detect
OpenAI publicly advertises analytics that can:
identify top users,
identify emerging usage patterns,
break usage down by individual user, product, and model,
distinguish valuable work from activity requiring closer review,
and grant individual power users more capacity so their work is not interrupted. (OpenAI)
Those are your words and your capabilities.
You built those tools for enterprise administrators.
Do not insult anyone’s intelligence by pretending OpenAI itself is technologically incapable of recognizing an extreme outlier in its own product.
My activity was associated with an account closely enough for your systems to:
calculate my usage,
display it to me,
deduct purchased credits,
enforce an account-specific limit,
and calculate the date on which I would be permitted to work again.
Whether the email address resides in the same database table is irrelevant bureaucratic trivia. The account is resolvable. The usage is attributable. The limit is enforceable. Authorized personnel can establish contact when the institution decides that contact matters.
Therefore, only two serious possibilities remain.
Possibility one: OpenAI’s systems failed to elevate an obvious, legitimate, productive usage anomaly to anyone capable of investigating it.
That is institutional blindness.
Possibility two: The activity was visible or flagged, and someone decided not to investigate or engage.
That is institutional cowardice or incuriosity.
I cannot prove which occurred.
But there is no flattering third possibility.
Either the largest artificial-intelligence company in the world could not recognize an enormous intelligence-product signal occurring inside its own platform, or it recognized that signal and chose silence.
Both answers are disqualifying for an organization claiming to understand the future before everybody else.
What Was Actually Being Built
This was not 6.8 billion tokens of idle conversation.
It was not role-play.
It was not engagement farming.
It was not a chatbot repeatedly talking to another chatbot.
I used Codex as an industrial cognitive environment while building and releasing Receiz: a proof-native software architecture in which digital objects can carry their own identity, provenance, ownership, custody, state, history, media, and continuity.
The work includes:
downloadable proof objects,
offline verification,
portable state and history,
an SDK,
an MCP server,
AI skills,
delegated authorization,
explicit capability signing,
deterministic reconciliation,
migration and conformance tooling,
live sports telemetry,
games built on proof-carrying creatures and cards,
production applications,
public release books,
repeated security and release qualification,
and more than one hundred versioned releases.
This is not a claim that exists only inside a ChatGPT conversation.
There are repositories.
There are packages.
There are applications.
There are tests.
There are release documents.
There are downloadable artifacts.
There is an offline verifier.
There are files that can leave the platform and carry the evidence required to verify themselves.
OpenAI says Codex is supposed to turn ideas into working software. OpenAI says more than five million people use Codex every week, and it promotes Codex as a system for increasingly complex work across engineering, research, analysis, operations, design, and other disciplines. (OpenAI)
OpenAI’s own research celebrates users assigning Codex longer-horizon work. It reports that the heaviest internal users regularly generate more than 60 hours of parallel Codex agent turns in a single day. (OpenAI)
So what exactly is the institutional response when a user takes that premise farther than your showcase, farther than your normal product assumptions, and possibly farther than anyone publicly documented?
Apparently:
Stop working and come back later.
That is not a frontier product response.
That is a vending machine response.
You Did Not Miss a Customer Testimonial
This is not about wanting OpenAI to tell me I am special.
You missed something much more valuable than a flattering case study.
You missed a live research laboratory.
You missed a frontier workflow.
You missed product intelligence.
You missed pricing intelligence.
You missed infrastructure intelligence.
You missed a legitimate stress test of Codex at extraordinary scale.
You missed an opportunity to understand how one person can coordinate sustained agentic work across an enormous production system.
You missed an opportunity to study where context caching succeeds, where agents repeat themselves, where token consumption explodes, where long-running projects retain coherence, where human judgment remains essential, and where limits interrupt rather than protect productive work.
You also missed one of the strongest positive public narratives available to you.
The public conversation around ChatGPT and AI repeatedly becomes dominated by dependency, delusion, cheating, displacement, manipulation, and fear.
The answer to that narrative is not another executive insisting that AI is beneficial.
The answer is evidence.
Show people remaining fully human, fully responsible, and fully capable of judgment while using AI to execute work that would otherwise require an institution.
Show the model as an instrument rather than an oracle.
Show a human using AI not to escape reality, but to create independently inspectable reality:
software,
infrastructure,
protocols,
products,
packages,
tests,
and artifacts that verify outside the conversation that helped create them.
That story was already happening inside your product.
You did not need to manufacture it.
You only needed enough organizational curiosity to notice it.
Instead, You Bought a Podcast
On April 2, 2026, OpenAI announced its acquisition of TBPN.
OpenAI said it wanted to use TBPN’s communications and marketing instincts to innovate in how it brings AI to the world and helps people understand the technology’s impact on their lives. (OpenAI)
Financial terms were not officially disclosed, but credible reporting placed the acquisition price in the low hundreds of millions of dollars. (Financial Times)
Consider that juxtaposition carefully.
You reportedly spent hundreds of millions of dollars acquiring people to talk about what artificial intelligence means while apparently having no functioning mechanism for recognizing a user already demonstrating what sustained human–AI collaboration can actually produce.
You purchased a conversation about AI.
You overlooked the evidence of AI.
I already have an audience approaching two million followers.
I already have the story.
I already have the distribution.
I already have years of public history establishing that these ideas did not materialize from one model prompt.
I already have working products.
I already have the repositories, packages, releases, and proof objects.
You did not need to purchase me.
You did not need to subsidize an entire media company.
You needed to send one fucking email.
This is not an argument that TBPN has no value.
It is an indictment of OpenAI’s sequencing and institutional priorities.
Before spending nine figures to improve how the world understands AI, perhaps the world’s largest AI company should first possess a competent system for understanding the extraordinary people already using its products.
You acquired a media organization because traditional corporate communication was apparently insufficient for the scale of your mission.
Fine.
But what does it say when the communications strategy can identify a podcast acquisition target while the product organization apparently cannot identify a 6.8-billion-token weekly outlier building production infrastructure?
It says OpenAI is better at buying narratives than discovering evidence.
Your Pricing Model Is Also Broken
After exhausting included usage, OpenAI offers Plus and Pro users paid credits as a usage extension. Included plan usage is consumed first; afterward, work begins drawing from the purchased credit balance. (OpenAI Help Center)
In my actual experience, $20 purchased 500 additional credits, and those credits disappeared extremely quickly under my workflow.
Meanwhile, another $20 Plus subscription on another email provided substantially more usable Codex capacity than the $20 credit purchase.
The repository did not move.
The folders did not disappear.
The code remained on the machine.
The productive environment was not meaningfully bound to the account supplying the model capacity.
That creates an obvious pricing arbitrage:
Twenty dollars spent as an existing power user buys less useful work than twenty dollars spent as a nominally new subscriber.
Why would a rational customer buy the credits?
Your pricing punishes continuity and rewards account multiplication.
It makes the overage product economically inferior to the acquisition product.
That means OpenAI is simultaneously:
subsidizing new included usage,
overpricing marginal usage for its most active customers,
creating incentives for fragmented accounts,
losing the behavioral continuity needed to understand those customers,
and interrupting the exact users most likely to reveal what the product can become.
I did not discover this through a pricing exercise.
I discovered it by trying to continue working.
That is strategic intelligence OpenAI should have wanted.
Nobody asked.
There Is No Economically Sensible Version of What Happened
Consider the possibilities.
If my 6.8 billion-token week was inexpensive
Then caching and model economics made the workload affordable enough that the system served it continuously.
In that case, abruptly stopping productive work at a coarse threshold was unnecessary, and OpenAI needs a legitimate power-user tier or individual capacity override.
If my 6.8 billion-token week was expensive
Then OpenAI funded an extraordinary consumption event without investigating whether it represented abuse, malfunction, inefficiency, or unprecedented productive value.
In that case, allowing the expense while learning nothing from it was fiscally irresponsible.
If the activity looked suspicious
Then investigate it.
Verify that a human is actively directing the work.
Inspect the repositories and releases.
Evaluate the output.
Classify the account appropriately.
If the activity was clearly legitimate
Then why was there no escalation?
Why was there no research outreach?
Why was there no product conversation?
Why was there no custom plan?
Why was there no effort to understand what the most extreme behavior revealed?
There is no rational branch in which the correct response is:
Fund billions of tokens, capture no strategic intelligence, contact nobody, and then freeze the work.
The first 6.8 billion tokens were apparently tolerable.
Then the next token became unacceptable.
That threshold may make sense to an automated allocation system.
It does not make sense as the final response of an intelligent organization.
At Your Scale, This Is Not “Startup Chaos”
On March 31, 2026, OpenAI announced $122 billion in committed capital at an $852 billion post-money valuation. OpenAI described itself as becoming the core infrastructure for AI and said Codex was transforming how developers turn ideas into working software. (OpenAI)
At that scale, institutional incuriosity is no longer cute.
It is not a scrappy startup forgetting to respond to an unusual support ticket.
The capital entrusted to OpenAI is supposed to purchase more than GPUs, acquisitions, launch videos, advertising, podcasts, and model training.
It is supposed to purchase the institutional ability to learn.
A company funded at this level should be capable of detecting unplanned upside—not merely controlling cost and risk.
It should know when a user is producing an extreme signal.
It should know how to distinguish extraction from creation.
It should know how to route a legitimate outlier to a human being.
It should know how to turn extraordinary product behavior into research, retention, revenue, and public understanding.
If OpenAI can miss a signal this loud, investors should reasonably ask what else it is missing:
What other frontier workflows are buried in usage logs?
What other new markets are being classified merely as compute expense?
What other builders are being stopped by generic limits?
What other pricing failures are obvious to customers but invisible internally?
How many valuable discoveries are dying between billing, safety, infrastructure, product, research, support, and communications because nobody owns the complete picture?
This is what happens when an organization has extraordinary local intelligence and catastrophic global stupidity.
Every department performs its function.
Nobody comprehends what the whole system is seeing.
And Yes, the Silence Looks Like Fear
I will not pretend to know the private motives of people I have never spoken with.
But behavior creates an appearance.
If OpenAI truly never noticed this usage, that is incompetence.
If OpenAI noticed and chose not to engage, the silence begins to look like fear—not necessarily fear of me personally, but fear of uncontrolled evidence.
Companies like stories they can frame.
They like benchmarks they designed.
They like demos they commissioned.
They like founders, partners, and case studies that arrive through approved channels.
They like narratives in which the institution remains the center.
What I built did not arrive through an accelerator, venture fund, enterprise partnership, media relationship, or internal referral.
It emerged from one person relentlessly using the product beyond the scale at which its assumptions remained comfortable.
And the result is not another AI wrapper.
It is architecture that reduces dependence on centralized authority.
The server is no longer God.
The object can carry its own proof.
The file can leave the platform.
The history can remain with the artifact.
Verification can happen offline.
Ownership and continuity do not need to disappear when a company, database, login, or API becomes unavailable.
That is not a tidy OpenAI customer story.
It is an uncontrolled technological result.
And uncontrolled results expose whether an institution genuinely wants innovation—or only innovation it can narrate, finance, and position before anyone else recognizes it.
Maybe OpenAI did not see it.
But you had every technical reason to see it.
Maybe OpenAI saw it and underestimated it.
But underestimation cannot make it disappear.
Maybe OpenAI assumed the usage limit would stop the work.
It did not.
It changed which account supplied model capacity.
It did not remove the repositories, the knowledge, the architecture, the releases, the applications, or the artifacts.
You can limit an account.
You cannot unbuild what has already been built.
My Questions for OpenAI
I am asking for direct answers:
Does OpenAI maintain any system for identifying legitimate extreme individual Codex users?
Did any automated system or employee flag my account’s usage for review?
If it was flagged, what determination was made?
Why does a user producing billions of tokens of legitimate Codex work receive an automated shutdown rather than human outreach?
Why does OpenAI advertise tools that identify enterprise power users while apparently lacking an equivalent internal discovery process?
Why can a new $20 subscription provide more useful Codex capacity than $20 in add-on credits for an existing user?
Why is there no rational individual power-user tier between ordinary consumer usage and fragmented credit purchases?
Why did OpenAI invest enormous capital in explaining AI publicly before constructing a competent system for discovering its strongest real-world users privately?
How many other builders are currently invisible to OpenAI because their work appears internally only as usage, cost, risk, or support activity?
Who inside OpenAI is actually responsible for asking what extraordinary usage means?
Do not answer this with a generic support template.
Do not tell me that privacy made detection impossible.
Do not tell me to purchase another small credit package without addressing the economic contradiction.
Do not celebrate multimillion-token demonstrations publicly and then act as though multibillion-token productive usage is unworthy of investigation.
You know enough about the account to measure it.
You know enough to charge it.
You know enough to limit it.
You can know enough to ask.
What OpenAI Should Do Now
Review the account.
Review the usage breakdown.
Inspect the work.
Speak with me directly.
Put Codex product, infrastructure, finance, research, safety, and communications in the same conversation.
Determine what the workload actually cost.
Determine which parts were cached context, productive output, repetition, model inefficiency, or avoidable consumption.
Study how the work remained coherent across releases.
Study how the human–agent relationship evolved.
Study why the credit product failed.
Create a real frontier-user program with consent-based outreach.
Create a power-user plan whose economics do not encourage fragmented accounts.
Create escalation rules for extreme legitimate production work.
And most importantly:
Develop enough institutional humility to recognize that the future of your product may emerge from somebody you did not select, fund, invite, or place on your podcast.
I am not asking OpenAI to invent my story.
I am informing OpenAI that it already happened.
You can engage with the evidence.
You can ignore it.
You can attempt to minimize it.
You can continue buying conversations about the future.
But you cannot claim the signal was too quiet.
It was 6.8 billion tokens in one week.
You counted every one of them.
And the proof is in the file.
In my case, literally.
BJ Klock
Founder, Receiz
July 20, 2026
