Those of us who have been with ChatGPT since day one and use it in all possible scenarios have been eagerly awaiting the new version of GPT-5. It was heralded in the AI circles as a real breakthrough — almost an “Oppenheimer moment” of artificial intelligence. But first impressions? Quite bitter. A big disappointment for GPT-5!
Big disappointment GPT-5! GPT-5 is supposed to be a step towards the holy grail — artificial general intelligence (AGI). OpenAI claims that the model is now “smart enough” to choose how to solve a task on its own. So it determines the “model” of operation for a given task. In theory, great, but in practice… If it were a free version or the one for 20 euros per month, we would still understand. But with a professional subscription for 200 euros, where experienced users are gathered, such an approach opens up more problems than solutions.
Now ChatGPT in all paid packages decides which “internal” model to use. The problem? For specific tasks, for which we professionals created custom GPT often chooses the wrong combination. Result: the answers are a step or two worse than before.
Lost Choice – The Big Disappointment of GPT-5
Previously, we knew exactly which model was best suited for a specific type of task. GPT-4.5 Research for Slovenian proofreading? Ideal. GPT-4o for image generation? Excellent. GPT-5 uses a less powerful model for easier questions, and for more demanding ones, it switches to a more advanced one (e.g. o3) at its own discretion. The user has no insight — nor control.
In editorial environments where custom GPTs were part of our daily workflow, we sometimes swapped models multiple times in the same conversation to get the optimal result. Now that luxury is gone. That’s why – GPT-5 is a big disappointment!
Is this progress or just server optimization?
After a few hours of use, the feeling is clear: GPT-5 is more of a server resource optimizer than a new, revolutionary intelligence. OpenAI clearly wanted to prevent users from unnecessarily running the most expensive models for trivial tasks. Logical from a business perspective, less logical for those of us who pay a hundred times more for a subscription than a typical user.
Truth be told — API access to the old models remains, but they have disappeared from the public application. And that's the problem. When ChatGPT decides which model to use, it often chooses a less suitable version than someone who knows exactly what they're doing would have chosen.
Using GPT-5 in its current form will likely result in a higher server load, as users — especially professionals — will have to use significantly more prompts than before to get the same quality results. Since we no longer have the option to choose the optimal model for a specific task, the work turns into a sequence of corrections, additional instructions, and retries. Ironically, this means more CPU usage, more server calls, and consequently more infrastructure load — the exact opposite of the goal for which this “simplified” approach was introduced.
Writing in the center, but the quality is no longer consistent
Most users ChatGPT is used for writing, research and document preparation. But quality in specific languages like Slovenian was not always a given — and now it is even less so. Previously, the typical workflow was clear: draft → stylistic upgrade with custom model → fact-checking with a fact-checker → proofreading. Each step with the model that was best for this task.
Now we have to prompt explaining how the model should behave in order to choose the right “internal” architecture. This is not optimization — this is additional work.
GPT-5 as a coordinator, not a wizard
After three hours of use, GPT-5 acts more like a coordination center for existing models than a new, unified intelligence. I'm not saying it's not truly better at certain tasks. It is. But its main task is to optimize server resources — and discipline "stupid" users who ran the most expensive model for a pancake recipe.
This is immediately apparent when generating images: GPT-5 often relies on o3, which is worse than GPT-4o in this regard. The result? The images are less convincing unless you manually switch the model — which you can no longer do now.
Conclusion: Advanced users are not impressed
For the average user, GPT-5 is probably magical. For a professional, it feels like someone took away their toolbox and left them with only a master key. It works… but slower, less accurately, and with more detours.
Or, as one long-time user summed it up:
"This is not progress. This is server optimization disguised as innovation."
