Description: Why keep ChatGPT for long conversations while handing development to Claude Opus 5.5? A first-person look at GPT-6, hallucinations, Codex limits, and a voice-input mishap involving Tibo.
One day I opened ChatGPT and discovered it had become GPT-6. Apparently, my AI had received a promotion without telling me.
OpenAI began rolling GPT-6 into everyday ChatGPT conversations on October 7, 2026. One change lets an answer appear while the model continues reasoning or using tools.[1][2] I had noticed that odd feeling: the reply starts early, then keeps growing. The time until the entire answer is finished can feel longer.
But the bigger change wasn't the model number. It was the unofficial organizational chart in my head.
ChatGPT runs the endless meeting
For talking things through, ChatGPT wins a lot of my attention. I can speak an unfinished thought, question the answer, reverse the argument, and repeat for ages. The conversation feels easy to sustain, and in my own use the limits don't interrupt me very often.
Think of it as a meeting room whose lights are always on. I can ask, “What am I missing?” and spend another hour pulling the thread.
Coding is different. For that, I tend to hand the job to Claude: inspect the project, find the cause, make changes, and check what actually works. Claude Opus 5.5, announced by Anthropic in September 2026, has felt remarkably quick for that workflow.[3]
Claude's chat messages also use capacity, so I'd rather spend that allowance on completing work than on forty minutes of philosophical wandering.
The result: the AI I want to talk to is not necessarily the AI I want to assign a ticket.
Less hallucination, or just harder to notice?
Paid, higher-reasoning ChatGPT feels much more dependable to me than some lightweight assistants I have tried. By the GPT-5.6 era, I was already struggling to feel the next increment.
That is an impression, not a controlled error-rate comparison across ChatGPT, Gemini, Copilot, or Claude.
What bothers me most is fabrication: a fake feature described confidently, a source that does not say what the assistant claims, or a “fixed it” message before anything has been verified. Smooth prose can make a bad answer more dangerous.
And accuracy alone isn't enough. A reply so short that it skips the useful reasoning is irritating too. Sometimes I want a real thought partner, not a fortune cookie wearing a lab coat.
Codex: excellent job title, awkward attendance record
Codex should be the programmer on this imaginary team. Yet in my use, the work often feels slow and the allowance disappears too quickly.
That is a personal experience, not a universal performance verdict. My practical test is simple: how much verified work did I get for my time, quota, and follow-up effort?
So why keep Codex around? Because of Tibo.
Thibault “Tibo” Sottiaux of OpenAI has publicly announced resets of usage limits for Codex and ChatGPT Work.[4] A depleted quota returns, and suddenly the backup developer has clocked back in. It feels less like infrastructure planning and more like collecting a surprise daily bonus.
The resets are discretionary, though. A responsible schedule cannot assume another one will arrive.
The assistant failed the easiest test: look up the name
Then voice transcription entered the story.
I said, “Because Tibo resets the limits.” The name came through badly. ChatGPT guessed, asked me to repeat myself, guessed again, and kept circling the mystery.
Finally I gave it the entire troubleshooting instruction in two words: “Search it.”
Once it searched, the public posts explained the reference. A supposedly brilliant thinking assistant had needed a human to remind it that search engines exist.
The takeaway isn't to crown one chatbot champion. It is to match the tool to the job, check actual outcomes, and look things up before inventing an explanation.
My current staff meeting is simple: ChatGPT thinks with me, Claude ships the code, Codex waits for Tibo—and I supervise the search button.
References (4)
- OpenAI, GPT-6 and Intelligent UI for everyone (2026-10-07) openai.com
- OpenAI, ChatGPT release notes, October 7, 2026 help.openai.com
- Anthropic, Introducing Claude Opus 5.5 (2026-09-22) anthropic.com
- Tibo (@thsottiaux), public Codex / Work usage-limit reset announcement x.com
