Over the past two days, I have been trying to understand a strange Codex usage event.
On July 19, shortly after my ChatGPT Plus access was restored, my entire weekly Codex allowance disappeared in approximately four hours.
That was completely different from my previous experience.
When Codex used a five-hour rolling allowance, I normally needed about an hour and a half of continuous development to reach the limit.
After the product moved to weekly limits, one weekly allowance usually lasted around four days.
The token activity data made the situation even more confusing.
On June 28, my profile showed approximately 110 million tokens used in one day, but I did not exhaust the weekly allowance.
On July 19, it showed only 39.056 million tokens, yet the entire weekly allowance disappeared in about four hours.
I cannot prove that OpenAI reduced the allowance. I also cannot determine whether the difference came from the model, cache behavior, or a change in usage metering.
But one thing was clear:
The event was far outside my historical usage pattern.
So I contacted OpenAI Support.
I Asked AI to Help Me Complain About AI
The situation was already slightly absurd.
I first used ChatGPT to organize the dates, screenshots, and historical data.
Then AI helped me write an English support request asking:
- Why did 39.056 million tokens exhaust a weekly allowance?
- Why had 110 million tokens not done so previously?
- Which models were actually used on July 19?
- Were there cache misses, retries, or duplicated context processing?
- Why was GPT-5.5 Thinking still visible but impossible to select?
I signed in to the OpenAI Help Center, opened the support chat, and submitted the explanation with screenshots.
OpenAI’s official support process begins through the Help Center chat bubble. Users first interact with a virtual assistant, and a human agent may join when the issue cannot be resolved automatically.
The first AI-assisted response was reasonably useful.
It explained that Codex limits may not be deducted according to raw visible token totals alone. Model choice, reasoning, tool use, and cache behavior may affect actual consumption.
It then asked me to provide a more detailed Usage breakdown showing:
- Models used by task
- Cached and non-cached tokens
- Credits consumed by task
- Whether Fast mode was enabled
There was one problem:
My Usage page did not show any of those details.
It displayed only daily token totals and the weekly allowance.
So I replied:
I cannot provide data that the product does not show me.
Please inspect the server-side records.
AI Escalated the Issue to a “Support Specialist”
The virtual assistant eventually responded:
This has been escalated to a support specialist. You should receive a response within the next few days, and replies will also be sent by email.
That made sense.
The AI assistant could not access the internal usage records, so it passed the case to someone who might.
A few hours later, an email arrived from OpenAI Support.
The agent accurately summarized my issue.
He knew that I was a ChatGPT Plus subscriber.
He knew that I frequently used Codex.
He knew that the incident occurred on July 19.
He knew that an entire weekly allowance had disappeared in approximately four hours.
He even understood that I was requesting an investigation into Codex usage metering.
Then he wrote:
It appears that you are not currently signed in to the account you are asking about.
I stopped for a moment.
But I Submitted the Case While Signed In
I was clearly signed in when I opened the support case.
The top-right corner of the Help Center displayed:
Logout (思吉 王)
The screenshots I submitted came directly from an authenticated Codex profile page.
The support email was delivered to the same email address connected to my ChatGPT Plus account.
The message contained the correct case number.
I continued replying from the same signed-in Help Center conversation.
In other words, OpenAI’s systems could:
- Determine which email address should receive the case;
- Associate the email with a support case;
- Read the full issue I submitted;
- Receive screenshots from my authenticated profile;
- Continue the conversation inside a signed-in Help Center session.
But they could not confirm:
Whether the person who submitted the case belonged to that account.
It reminded me of a famous Chinese expression.
From “Prove Your Mother Is Your Mother” to “Prove Your Account Is Your Account”
Years ago, China had a public debate about absurd administrative certificates.
The most famous phrase was:
Prove that your mother is your mother.
People could already have identity cards, household-registration records, and government documents, yet still be asked to provide another certificate because different institutions did not share or trust one another’s data.
“Prove your mother is your mother” became a national joke.
It represented more than an inconvenient document.
It represented an absurd system:
Every department possesses part of the truth, but no department is willing to connect the evidence and make a decision.
My experience with OpenAI Support followed almost exactly the same logic.
Previously:
The bank knows who your father is. The registration system knows too. You still need to prove that your father is your father.
Now:
OpenAI knows my email, support case, Plus account, and Codex data. I still need to prove that my account is my account.
The technology changed.
The process did not.
The Model Was Not the Main Problem
The interesting part is that I do not believe the AI failed to understand me.
The first virtual assistant understood the issue.
It recognized the date, understood that weekly allowance and raw token totals might not be linear, and identified the data required for an investigation.
The email agent also understood the issue.
He reproduced it almost completely.
The process failed elsewhere:
- The support system was not correctly connected to the product account;
- Help Center authentication did not become a verified support identity;
- The AI could not access usage records;
- The human agent saw an “unverified” status and followed a template;
- Nobody was empowered to connect the evidence already available.
This may be the most important limitation of many AI customer-service systems.
A company places a powerful language model inside a chat window without rebuilding the permissions, data connections, and operational workflow behind it.
The result is a model capable of understanding a complex complaint—but unable to perform the simplest action:
Confirm that a signed-in user is signed in.
The Smarter the Model Becomes, the More Absurd the Old Workflow Looks
With older chatbots, users understood the problem.
They recognized keywords.
They offered numbered menus.
They frequently answered the wrong question.
That was expected.
Modern language models can summarize long conversations, interpret screenshots, analyze unusual data, and propose reasonable investigation steps.
When such a system still sends the user into an impossible verification loop, the experience becomes even more frustrating.
The user can clearly feel:
It understands—but it cannot do anything.
It can say the correct words.
It cannot access the correct data.
It can make the company sound intelligent.
It cannot take responsibility for a decision.
This may be the central illusion of AI customer service:
The company believes it has deployed a smarter support agent.
The user receives a traditional workflow robot that is simply better at writing.
The Hardest Part of an Agent Is Not Tool Calling
The AI industry is currently focused on agents.
How models invoke tools.
How they connect through MCP.
How they access enterprise systems.
How they execute tasks automatically.
But this experience made me think that the hardest question may not be whether AI can call a tool.
It is whether:
- The AI can access the right data;
- It knows that records across systems belong to the same person;
- It has authority to make a decision;
- It can cross departmental boundaries;
- Someone accepts responsibility when systems conflict.
Without those foundations, even the strongest agent stops at the entrance to the workflow.
It can ask me to provide a Usage breakdown.
It cannot know that my interface does not expose one.
It can send a support email to my account address.
It cannot confirm that the address belongs to the account being discussed.
It can escalate the case to a human.
The human can send me back to the login flow.
The result is a perfect loop:
AI tells me to find a person. The person tells me to return to the system. The system asks me to prove that I am myself.
This Is Where AI Products Actually Need to Change
Many companies are adding AI to existing products.
They add chat windows.
Intelligent search.
Automated replies.
Agents.
But if the foundation still consists of disconnected databases, rigid permission labels, and workflows with no clear owner, AI only makes the surface appear more advanced.
It does not automatically make the organization intelligent.
A genuine AI product transformation also requires redesigning:
- How identity moves between systems;
- Who owns each data source;
- What the AI is permitted to see;
- When a human takes over;
- Whether users must repeat information already provided;
- Who is responsible for resolving the case.
Otherwise, we will continue creating new digital versions of the same old joke:
Please prove that you are signed in.
Please prove that this is the email address to which we just sent an email.
Please prove that this account screenshot came from your account.
Please prove that your account is your account.
Final Thought
My Codex usage issue remains unresolved.
I am still waiting to learn whether OpenAI Support can inspect the models, cache behavior, Credits, and tasks recorded on July 19.
But the support experience has already answered a different question:
A language model can change how customer service communicates. It does not automatically change how a company resolves problems.
A decade ago, people laughed at the phrase “prove your mother is your mother.”
Today, I signed in to one of the world’s most advanced AI products, used AI to write a complaint, and had an AI-assisted support system send it to a human.
The final request was:
Please prove that this account is your account.
The AI era has arrived.
Some workflows are still living in the previous one.
Estimate task cost in the AI Cost Estimator or refine prompts in the Prompt Optimizer.