GPT-5: The Day OpenAI Broke Everything

The Day OpenAI Lost Me

I am going to try to explain this without sounding dramatic, and I am going to fail, because it was dramatic. August 7th, 2025. OpenAI flipped the switch on GPT-5 and, in that first moment, yanked every previous model out from under us. No GPT-4. No GPT-4o. Just their shiny new thing that, as it turns out, could not hold a conversation for more than three exchanges without eating itself.

If you do not use these tools daily, this might sound like a minor inconvenience. Let me recalibrate that for you.

What ChatGPT Actually Was to Me

I am all-in on Claude Code for development. When I need to ship code, architect systems, debug something gnarly - Claude is my copilot. In December 2024, I built an entire Electron app with Rust binaries using ChatGPT’s browser interface. I had never touched Rust in my life. The app works. Ninety-five percent of that code was AI-generated, and it runs. That is not a hypothetical future. That already happened.

But ChatGPT served a different role. It was my thinking partner. The place I went to brainstorm, to workshop ideas, to explore topics without the pressure of production code. Professional projects, personal curiosities, random 2 AM thoughts about whatever my brain latched onto because undiagnosed ADHD does not have business hours. ChatGPT was where those conversations lived.

Was. Past tense.

The ADHD Factor

Here is something the tech press does not talk about when they review AI models: for people who think the way I do, these tools are not just productivity boosters. They are cognitive prosthetics. My brain does not do linear thinking. It does not do “let me methodically work through this problem step by step.” It does pinball. It does “here are seven tangentially related ideas and I need to externalize all of them right now before they evaporate.”

ChatGPT was the perfect receptacle for that. I could dump half-formed thoughts into it and get back structured responses that helped me figure out what I actually meant. I could bounce between topics at the speed my brain wanted to move, and the model kept up. When a tool works like that for a brain like mine, it is not a convenience. It is the difference between a productive day and a day spent staring at a wall trying to remember what I was about to do.

So when OpenAI broke that tool and, at least in those first hours, removed the option to use the one that worked? That was not a product downgrade. That was pulling the rug out from under someone who had built their entire cognitive workflow around it.

What GPT-5 Actually Did

Picture this: you ask a straightforward question. The model starts answering, then catches itself in some recursive loop, repeating the same three sentences like a broken record. You try to redirect. It apologizes, starts fresh, then - right back into the loop. Simple math becomes a coin flip. Counting letters in a word is apparently PhD-level difficulty. Code generation produces hallucinated APIs that have never existed.

The Reddit threads told the story: “GPT-5 is horrible” with thousands of upvotes. Users discovering their $20 per month now bought them a model that could not count to ten reliably. And the worst part, in that first stretch? No obvious way back. No “use GPT-4 instead” toggle staring you in the face. Just this broken thing and a company telling you it was their “smartest, fastest, most useful model yet.” To their credit, OpenAI did walk that back within days, restoring the older models like GPT-4o to the picker after the backlash - but the launch-day shock had already done its damage.

The benchmarks they were touting - 94.6% on AIME 2025, 74.9% on SWE-bench - those numbers mean nothing when the model cannot maintain a coherent conversation. I do not care about your benchmark scores. I care about whether the tool works when I sit down to use it. It did not.

The Trust Problem

What really got me was the arrogance. OpenAI knew this model was not ready. Sam Altman himself warned about “capacity crunches” and “hiccups.” They launched anyway. Then they removed access to the stable, functional models we had been using for months. That is not a confident product launch. That is a forced migration designed to make their metrics look good.

Trust is earned in drops and lost in buckets. I used to recommend ChatGPT to everyone. “Try it,” I would say. “It will change how you work.” Now? I tell people to look at Claude first.

Where I Landed

Deeper in the Claude ecosystem. That is where I landed. Claude Code is my development environment. Claude is my thinking partner now too. The transition was not seamless - Claude has a different personality, a different rhythm - but it works. And more importantly, Anthropic has not yet pulled the rug out from under me.

My half-brother has a PhD from Berkeley and teaches AI in Vancouver. He lives in the world of how these models are built. I live in the world of how they are used. And from the user side, I can tell you: the best model in the world is worthless if the company that makes it treats you like a beta tester instead of a customer.

The ChatGPT app still sits on my phone. Occasionally I open it, hoping maybe they have fixed the loops or just acknowledged the problem. Meanwhile, my Electron app built with ChatGPT in December 2024 still runs perfectly. The irony is not lost on me: their old model helped me build something that works better than their new model does.

GPT-5 was not just a bad launch. It was a lesson in what happens when a company forgets that their product is not the model - it is the relationship with the people who use it. OpenAI shook that relationship on August 7th, and unlike their models, that is something I will not forget.