ChatGPT-5 is just around the corner – but the real change is less technological than structural. OpenAI is planning to automate model selection. In future, you will no longer decide whether GPT-3.5, GPT-4 or GPT-4o is used. The system will make this decision itself. Internally, this is called “dynamic routing” – a process in which the AI itself determines how much computing power it uses and which model is activated, depending on the task at hand.
What is sold as progress is actually a loss of control. Because with automated routing, the very element that previously provided clarity disappears: the conscious choice of model by humans. Instead, in future you will interact with an interface that suggests that “everything is under control” – while you don’t know which model is talking to you, what limits it has and whether it even knows your context.
For the average user, this may not matter. For professional users, it is a problem. Because anyone who relies on consistent performance, memorable contexts and traceable statements not only needs a good model. But also transparency, control and responsibility.
Who is actually speaking?
Nate B. Jones is an AI strategist, former Head of Product at Amazon Prime Video – and one of the keenest observers of current AI developments. His briefings on the future of ChatGPT, OpenAI & Co. reach over 250,000 followers on TikTok.
Why this is problematic – even if you are a professional
A viral TikTok video puts it in a nutshell: ChatGPT claims that GPT-4o is better than O3 – although the opposite is true. O3 is the newer, more powerful model that outperforms GPT-4o in many benchmarks. So why this mistake? Not out of stupidity. Not out of hallucination. It’s because the system has learned to guess when it thinks you’re expecting a simple answer.
This is not a glitch. It’s part of the design.
These models are designed for efficiency. They appreciate because they want to save tokens. And they answer as they have learned: with a convincing pitch, even if the content is questionable. In the TikTok example, the cause lies in the naming: the model concludes that “4” is greater than “3” – and draws a seemingly logical conclusion from this. But it does not test it.
That would be forgivable when it comes to small talk. But what if you’re writing a legal opinion, designing a market model or obtaining a medical recommendation?
Then apparent intelligence becomes a risk.
Because you are not talking to a knowing authority – you are talking to a system that confuses probability with truth.
And this is precisely where the explosive power of automated routing lies: if a system decides for itself how much it thinks, what it can do – and above all when it should rather guess – then the result becomes unpredictable. Not because the model is weak. It’s because you no longer know which cognitive architecture was active in the background.
Personal experience: three models, three personalities – and one problem
I work with different GPT models every day – and can now tell which one is responding with my eyes closed.
GPT-4o knows me.
It remembers our conversations, respects my tone of voice, understands my style. The collaboration is fast, precise – sometimes almost too pleasing. 4o often proves me right. Too often. But I’ve learned to deal with it. I then question myself. And yet: I feel that I’m working at eye level.
GPT-4.1 also knows me.
Personalization is also active here. The collaboration is a little more sober, but stable. There are fewer queries, the model seems more focused – and I know where I stand.
GPT-4.5, on the other hand, does not know me.
It calls me by my first name. It ignores my vocabulary. It answers questions precisely – but coldly. No context, no relationship, no history. That may be enough for a new start. It’s not enough for strategic work.
The decisive factor: I know which model speaks when.
I can see the differences, classify them and react to them. If the new routing with GPT-5 makes this difference invisible, we not only lose the choice – we lose the relationship.
What you need to do to stay in control
The biggest danger in dealing with ChatGPT-5 is not the model itself. It is your trust. Trust in a system that, in case of doubt, doesn’t know that it’s wrong – and that you soon won’t even be able to identify.
But you can – and must – take countermeasures.
The most important principles
Here are the key principles for surviving the new routing without falling victim to your own efficiency:
1. tell the model how it should think
Not metaphorically. Literally. Example: “Please do not make assumptions. Do not guess. If you are unsure, ask a question or activate the web search.”
GPT-5 can work very precisely – if you guide it precisely. It is not a thinking partner. It is a probability model with access to an infinite amount of knowledge – and without any awareness of truth.
2. use the browser function – but not blindly
Just because a model “was on the Internet” does not mean that its answer is reliable. Let them show you what they found.
Actively ask for the source. Ask for quotes, screenshots, URLs. And check: Is the knowledge structured? Is it outdated? Relevance or just noise?
3. secure your context
If you work systematically with ChatGPT – for research, text production or strategy development – then ask regularly:
“What else do you know about my project?” “What assumptions are you making right now?”
The model does not forget. But it prioritizes – and you need to know where you stand in its logic.
4. document which model you have used
Especially for professional output: Note whether you have worked with GPT-4o, GPT-4.5 or GPT-5.
Because if you rely on statements – in contracts, presentations, recommendations – then you must be able to say in an emergency: “This statement came from this system, with these conditions.”
5. don’t be fooled by the surface
GPT-5 will appear smoother, faster and more comfortable. But that’s precisely what makes it dangerous. It sounds confident – even if it is wrong. It seems intelligent – even if it only recognizes patterns.
Therefore, the friendlier the surface, the more important your critical eye is.
”ChatGPT-5 will undoubtedly make many things better. It will be faster, more flexible and, for many people, more "magical" than anything before. That's exactly the problem. Because the smoother the interface, the less you feel the break - the point at which the system no longer knows what it is doing. If you then believe that the model thinks "for you", you've already lost.
Norbert Kathriner
List of all new features of ChatGPT-5
ChatGPT-5 brings numerous significant innovations that represent both technological and functional improvements over previous versions.
Many of the extended functions (e.g. extended intelligence modes, larger context window) are reserved for paying users for the time being. New features such as permanent memory and the multimodal interface will be introduced gradually and may not be activated in all regions or for all users immediately at launch. This list is based on the information currently available as of July 2025 and may become more specific when official product announcements are made.
- Unified model: ChatGPT-5 replaces the previous variety of individual models (such as o-series, 4o, 4.5, etc.) with a centralized AI that integrates all functions and automatically selects the appropriate internal module for each request.
- Free, unlimited access: For the first time, users get unlimited access to the standard intelligence mode of ChatGPT-5, regardless of whether they have a paid Plus or Pro subscription.
- Extended multimodality: ChatGPT-5 processes and generates not only text, but also images, audio data and, probably for the first time, videos. This makes it possible, for example, to upload a video for automatic summarization or analysis.
- Significantly greater context length: The amount of text that can be processed simultaneously (“context tokens”) is increased to up to 256,000 tokens – this means that very extensive documents, files or databases can be analyzed directly.
- Chain-of-thought reasoning: In future, complex tasks will be systematically broken down into individual steps (“chain logic”). This increases efficiency and precision when solving problems and significantly exceeds the capabilities of GPT-4.
- Improved personalization: AI can permanently take individual preferences, histories and specific user needs into account and thus provide tailor-made answers.
- Reduced hallucinations: The number of incorrect or made-up answers (“hallucinations”) has been greatly reduced through improved training and greater data diversity. ChatGPT-5 is considered more reliable and accurate than all previous versions.
- Agent mode and autonomous functions: ChatGPT-5 is able to autonomously plan and manage tasks (e.g. scheduling, e-mail management) and also perform longer “in-depth searches” autonomously.
- Infinite memory: The model can permanently remember all previous interactions with a user, which considerably simplifies and speeds up interaction. This function has not yet been rolled out everywhere.
- Improved speech processing and speech cloning: After just 15 seconds of speech input, ChatGPT-5 can generate an artificial speech model that sounds extremely similar to the original.
- Open Deep Research function: Features such as “Deep Research” (extensive, page-long analyses) are no longer only available to Pro users, but – with restrictions – to all users.
- Full API support for new task formats: Image and video generation are now also automatically accessible via the API.
