For several months, we'd been waiting for GPT-5, announced as THE new super-powerful model, bringing us closer to AGI (superintelligence).
Spoiler: we're still far from it
Before sharing my disappointment, I wanted to go back 8 months, to late 2024.
At that time, the first reasoning models were starting to come out. Very powerful models, able to take time to think (reason) in order to generate more reliable answers by mobilizing more resources.
The gap between models like GPT-4 and GPT-o1 was huge in terms of performance. It gave us many more use cases (and unlocked automation cases that required fairly heavy checks).
This logic of reasoning before answering changed a lot of things, and personally, it's the type of model I was using constantly.
(→ regarding the image, shared ChatGPT conversations (public ones) are no longer indexed on Google)
What's new
And the big change is the unification of reasoning models with classic models. This puts an end to the explosion of models, a good thing rather than having: 4o, o1, o3, 4.1, etc.
A single model that decides How To Answer (and which tool to use)
On top of that, the news in bulk:
- Access to the new models for ALL users
- GPT-5 Auto, Fast (=gpt-4o) and Thinking (=oX)
- GPT-4o for paying users (older model)
- A faster, more powerful model with fewer hallucinations
- A more “human” answer (writing) style
- And a script execution window
The disappointment…
And it's precisely this approach that affects power users, those who switched between several models depending on how they wanted the model to answer.
For me, it's a huge loss of control, because the model decides on the prompt's complexity, mobilizing either the fast model (simple question) or the reasoning model (complex question).
But in some cases, I want the model to be more thorough even for a simple question. A bit like not wanting to get it wrong… But of course, that has the downside of costing more for OpenAI…
Tip: you can explicitly ask it to “think before answering” to encourage activating the Thinking mode (reasoning)
I use OpenAI 90% of the time rather than another LLM, mainly because it's the most advanced ecosystem → you find the best models and the latest features (deep research, ChatGPT Agent, image generation, etc.).
With every announcement, it's often spectacular. But not this time (in my case).
Even though for most users, GPT-5 will let them discover reasoning models (previously reserved for those who deliberately changed the default model), and therefore have a far better experience.
And in any case, the model is faster, and the benchmarks show it more performant.
A break in trust?
In some cases, I found myself in an embarrassing situation where the GPT-5 model was completely hallucinating on basic questions. Here's an example:
The model's error mainly came from using “.” instead of “,”. But the Thinking model was able to find the right answer.
Since then, this type of error has been fixed. But it has the particularity of causing a loss of trust, a constant doubt that pushes me to keep favoring the Thinking reasoning mode.
Ethan Mollick pointed out that you could have significant performance gaps in the Auto mode of GPT-5, sometimes landing on good results and sometimes on poorer ones.
Personalizing the models
When there's a model change like this one, what can be frustrating is the shift in writing style that can occur. Even if you can go further with custom instructions.
Many users did it on GPT-4o to get a more personal style, in tune with their expectations.
A model change means a change in behavior.
And it can be hard to get back the style you had created, and to recover, in a way, the model that had been shaped.
Even though models don't improve the more you use them, context elements are stored in memory, letting the models retrieve certain elements when relevant.
This trend of cross-chat memory + profile personalization is becoming common for most models. Claude announced it recently (see below) and Gemini has just started doing it too.
