OpenAI has built its brand on productivity. ChatGPT is sold as a tool that helps you think, write, code and finish work faster, and Codex promises to slot into real development workflows. I believed that and paid for it, because I wanted to save time, cut repetitive work and focus on what actually matters.
For a while that felt true. Then the product started working against me, and the company's support finished the job. This isn't a generic essay about the limits of AI. It's my own experience with ChatGPT, Codex and OpenAI's customer support.
The breaking point
I was doing simple website content edits in Codex, nothing complex or demanding, when it switched to the Astra model on its own. I hadn't chosen it. I already knew from experience that Astra eats through usage allowance quickly, which is why I only ever used it deliberately and sparingly. When I picked it myself, it cost me only a small slice of my limit.
This time was different. After the unexpected switch, my entire five-hour limit disappeared, even though most of my weekly allowance was still untouched. I couldn't keep working. Two other tasks were interrupted, my concentration was gone, and I had to spend time dealing with a problem that had nothing to do with the work itself.
That is the opposite of what a productivity tool is for. I pay to make my work easier, not to troubleshoot the tool's own behavior. I reported the incident expecting an investigation. I didn't expect perfection, but I did expect accountability.
What I actually asked for
I never demanded my usage to be restored in full. I asked for one additional reset, so I wouldn't have to spend either of the two resets I had already earned. I saw it as a modest compromise that would solve the immediate problem and keep the relationship intact.
OpenAI's reply was that their data couldn't show whether the model switch was a malfunction or whether usage had been recorded incorrectly. Then they offered me two choices: use one of my existing resets, or wait for the next automatic one.
That missed the point entirely. Offering me something I already own is not compensation or a resolution. In effect, I was being asked to pay for an incident that I believed was caused by their own software. I also explained that this wasn't only about credits. I had lost working time and momentum. It changed nothing.
Nobody is home: support run by a bot
I sent screenshots of the usage dashboard, context about my earlier Astra usage and a reference to the exact Codex thread. OpenAI runs the platform, stores the conversations and controls both model selection and usage accounting, so I expected a real technical look at its own data. What I got was another explanation of why they couldn't give me a reset.
I asked several times for a human. Every reply was signed "Benjamin", and I was told my case had been passed to a specialist. But the answers were generic, repetitive and detached from what I was actually arguing, and they never engaged with a single one of my points.
So I ran a small experiment. I pasted the whole ticket history into ChatGPT and asked it to answer as a support agent. The result was strikingly close to what OpenAI had sent me.
My conclusion is that neither in chat nor by email did a real person answer me. I was talking to an AI assistant that goes by the name Benjamin. I can't prove that from the outside, and OpenAI's own help page says only that a virtual assistant answers first and a human agent may join later. But everything about my experience says no human ever joined. I was arguing with a machine that was programmed to say no politely.
That is the heart of the problem. A company that sells AI to the world should be the last one to hide behind it when a paying customer asks for a human being.
The everyday frustrations
Codex was the last straw, not the only one. My dissatisfaction had been building for a long time, and the same problems kept coming back.
Too many words, too little substance. Simple tasks turn into long explanations. A direct request brings unsolicited advice, commentary and irrelevant detail, and the useful part ends up buried. I don't need an assistant that tries to impress me with how much it can write. I need one that understands the request and delivers the result.
Weak instruction following. I give clear, structured, unambiguous instructions and still get something different back. Requirements are ignored, details are changed without permission, and constraints are overlooked. After a while it stops being assistance and becomes babysitting.
The same mistakes, again and again. ChatGPT acknowledges a correction, says it understands, and then repeats the error. Sometimes the task is so simple that I can't see why it needs several rounds. I have no evidence that the models sabotage tasks on purpose. But after enough ignored corrections, you start to wonder whether they follow instructions at all.
Confident errors. Wrong information delivered in an authoritative tone is worse than an honest "I'm not sure". I can't rely on the answer, so I have to verify it, hunt for mistakes and sometimes undo decisions I made based on it. An assistant that needs constant supervision is not a productivity tool.
Unpredictable usage. When models differ so much in cost, users need real control over which one runs. An unexpected switch that drains a limited allowance isn't a minor inconvenience.
Promises versus reality
A flashy demo is not the same as a reliable product. When I use a tool professionally, I need it to follow explicit instructions, produce usable results, behave predictably, avoid surprise interruptions, and come with support that can investigate real problems. These are basic expectations for a paid subscription.
I know AI has limitations, and I don't expect every answer to be perfect. But a company charging for access has to own the whole experience: the models, the interface, usage accounting, model selection and support. The customer doesn't experience separate components. They experience one product, and it either helps them work or gets in their way. For me, ChatGPT increasingly did the latter.
An avoidable loss
What bothers me most is how easily this could have ended differently. I wasn't asking for money. I wasn't asking OpenAI to admit a bug without checking. I offered a simple compromise. They could have investigated the model switch, explained what happened, put a real person on the case and made a small goodwill gesture. Instead I got the same refusal over and over, in different wording.
So I cancelled and moved to Claude. A customer reported a problem, backed it up with evidence, explained the consequences, offered a compromise and asked repeatedly for a human decision-maker. The company failed to resolve it, and the outcome was predictable. OpenAI lost a paying customer.
Why I wouldn't recommend ChatGPT to anyone
I don't say this lightly, and it comes from my own experience, not from benchmarks. But if a friend or colleague asked me today, I would tell them to look elsewhere, for five reasons.
First, it costs you time. The promise is that it saves you effort, but in my experience I spent more and more of it correcting answers, repeating instructions and checking results.
Second, you can't fully trust it. Confident mistakes force you to verify everything, which cancels out much of the benefit.
Third, your workflow isn't safe. As the Codex incident showed, an unexpected model change can drain your allowance and stop your work with no warning.
Fourth, nobody will help you when it goes wrong. When I hit a real problem, I got scripted answers and no investigation.
Fifth, you may never reach a human. In my case, both chat and email were handled by what I'm convinced was an AI assistant, "Benjamin", that never once engaged with my actual arguments. If a company won't put a real person behind its support, you're on your own as a paying customer.
OpenAI didn't lose me over a slightly better benchmark score elsewhere. It lost me because using its products became exhausting, and when I asked for help, I got more of the same.
I came for productivity. I left with frustration.
Goodbye, ChatGPT. Hello, Claude.
