ChatGPT-5: Fewer Hallucinations, More Limitations? What I learned following OpenAI’s annoncement
- Marie Horodecki Aymes
- Aug 8
- 3 min read
So, what exactly is GPT-5?
OpenAI is rolling out GPT-5 to paying ChatGPT users (Plus, Pro, and Team plans).The message is clear: GPT-5 is smarter, more reliable, and most importantly... makes fewer mistakes.
That’s good news.
Because previous versions (even the most advanced like GPT-4o) had a real accuracy problem.
What I was hoping for in a new model
I’m someone who uses ChatGPT regularly. But less and less for analysis: debugging hallucinations often takes me more time than producing the analysis myself.
Even when I provided clear source documents and told it explicitly to stick to them, the model would drift (sometimes slightly, sometimes a lot). So I focused on what still worked well: text editing.Rewording, adjusting tone, clarifying paragraphs ... for that, it remains useful and fast.
And for coding?
I mostly used GPT-3.5 for basic Python. Nothing too technical, but enough to automate tasks, test ideas, or save time.
And honestly, GPT-3.5 got the job done.To my developer friends: no, I haven’t suddenly become an IT expert (I’m staying in my comfort zone).
GPT-4o could be overly creative or slow at times, but overall, it handled those use cases decently.
What GPT-5 promises
I don’t have access to GPT-5 yet.
But I’ve read OpenAI’s official documentation closely.
Here’s what stands out:
GPT-5 makes fewer errors.
It can engage in deeper reasoning when a task requires it.
It can automatically switch into "thinking mode."
It still supports all the existing tools (file analysis, images, web browsing, etc.).
That’s reassuring, especially for people dealing with technical or sensitive content.
But there’s a catch
This new model comes with some important changes in how users can interact with it.
1. You no longer choose your model
Previously, you could choose GPT-3.5 or GPT-4o depending on your needs. With GPT-5, that’s over.OpenAI is phasing out the older models for most users.
2. Access is more limited
If you’re on the ChatGPT Plus plan (like I am), here are the new limits:
80 messages every 3 hours.
200 messages per week if you manually select the “Thinking” mode.
After that, you’re automatically switched to a lighter version of the model with fewer capabilities. There’s no way around it.
3. The model decides for you
GPT-5 analyzes your prompt and decides on its own whether to think harder or not.That’s supposed to make things easier, but it also removes some control from the user.
My takeaway so far
GPT-5 finally seems to fix a real issue: repeated factual errors.That’s what cost me the most time — and at times, trust in the tool.
But in parallel, the overall user experience is changing:
Quick Comparison: Before vs After GPT-5 (ChatGPT Plus)
Feature | Before (GPT-4o / 3.5) | After (GPT-5) |
Model selection | Yes (GPT-3.5, GPT-4o) | No, single model with auto-switching |
Message volume | Unlimited or very high | 80 messages every 3 hours |
Deep reasoning | Always accessible | Limited to 200 Thinking messages/week |
Hallucinations / factual errors | Frequent | Promised reduction |
Response speed | Fast (GPT-3.5), slower (GPT-4o) | Fast (Chat) or slower (Thinking) |
User control | High | Reduced (model chooses response depth) |
Coding (Python) | Smooth with GPT-3.5 | Better reasoning, but quota to monitor |
What’s next?
I’m waiting to try GPT-5 myself to see how all of this holds up.In the meantime, I think it’s important to look beyond the promises.A good model isn’t just about better answers.It’s about being a good tool — something useful, predictable, and manageable.
Provisional conclusion: fewer hallucinations, more power, but also more restrictions. GPT-5 may be a step forward, but in what direction?
Comments