GPT-4.5: Hype or game changer?

“GPT-4.5 is at the frontier of what is possible with unsupervised learning, a method where AI independently recognizes patterns and structures without predefined labels. We continue to be surprised by the community’s creativity in discovering new possibilities and unexpected applications.”

That is what OpenAI wrote at the launch of GPT-4.5, a new AI model that promises to be smarter, more intuitive, and more accurate than its predecessors. For now, the model is exclusively available to users of ChatGPT Pro (USD 200 per month, which is also the price in the Netherlands and other European countries) and API customers. OpenAI positions GPT-4.5 as the most advanced version of its generative AI, featuring improvements in text comprehension, creativity, and conversational skills. However, various benchmark tests and practical experiences show that the model is not making the expected progress in crucial areas. What is new about GPT-4.5? What reservations can be made regarding this new model? And do the performances justify the high costs?

What is new in GPT-4.5?

GPT-4.5 has been trained with more data and computing capacity than its predecessors. This results in:

• Better understanding of user intent – subtle signals and complex prompts are interpreted more accurately. Where earlier models struggled with ambiguous or implicit questions, GPT-4.5 demonstrates a better ability to understand context and correctly infer user intent.
• More natural communication – responses feel less ‘robotic’ and more conversational. The responses from GPT-4.5 are more fluid, empathetic, and better aligned with the user’s tone. This makes the model better suited for interactions where nuance and emotional intelligence play a role.
• Fewer ‘hallucinations’ – OpenAI claims that GPT-4.5 is less likely to produce inaccurate or misleading answers. Although AI models sometimes tend to present fabricated facts as truth, GPT-4.5 should theoretically perform better in this area. However, initial tests show that hallucinations still occur.
• Higher accuracy in Q&A benchmarks – the model performs better on Simple QA tests than its predecessors. This suggests that the model is more reliable in correctly answering factual questions, which is particularly beneficial for applications such as search engines and knowledge bases.
• Broad compatibility – GPT-4.5 works with existing functionalities such as file uploads and canvas, allowing users to upload and edit files more easily within the same interface. This offers more flexibility for professionals who want to integrate AI into their workflow.
• Higher creativity – the model excels in creative tasks such as content generation, brainstorming sessions, and writing support. This makes GPT-4.5 particularly suitable for copywriters, marketers, and content creators who need an AI that thinks along and can provide original ideas.

Progress with reservations

Although GPT-4.5 shows improvements in some areas, the release also raises questions. This is partly because OpenAI quietly removed the description of GPT-4.5 as ‘not a frontier model’ from the original GPT-4.5 documentation. Furthermore, the training dataset is no newer than previous models, with a dataset that was updated until October 2023.

Other reservations:

• Not always better than previous models – in benchmarks for mathematics and logical reasoning, GPT-4.5 lags behind specialized models such as DeepSeek R1 and Claude 3.7 Sonnet. This means the model still struggles with complex analytical tasks and high-level reasoning.
• High costs – the model is significantly more expensive to use (see the next section), raising the question of whether the performance justifies the additional cost.
• Doubts about scalability – AI experts speculate that the benefits of scaling generative AI are diminishing. While earlier GPT models showed enormous leaps in performance by simply adding more data and computing power, this effect seems to be leveling off with GPT-4.5.
• Mixed results in complex analyses – for legal, scientific, or financial applications, GPT-4.5 offers no significant improvement over previous models. This means that professionals in these sectors may see little benefit from upgrading to GPT-4.5.
• Better interaction, but no major breakthrough – the model scores well on human interaction and creativity but shows little progress in the area of deep reasoning. Consequently, relatively little progress is made with GPT-4.5 regarding complex tasks, in-depth analysis, and the understanding of abstract concepts.

The costs

One of the biggest points of criticism is the price of GPT-4.5.

OpenAI applies a significantly higher rate than with previous models:

• $75 per million input tokens and $150 per million output tokens – for comparison: GPT-4o costs only $2.50 per million input tokens and $10 per million output tokens. This makes GPT-4.5 no less than 30 times more expensive.
• High costs for API users – organizations wishing to develop AI applications with GPT-4.5 must pay substantial amounts. This could make the model inaccessible to startups and smaller organizations.
• Doubts about long-term availability – OpenAI itself has indicated that it is uncertain whether it will continue to offer GPT-4.5 in the API in the long term, due to the extremely high costs of running the model. This suggests that OpenAI also recognizes that GPT-4.5 is financially less attractive than previous versions.
• Only available for Pro users – currently, GPT-4.5 is only available to ChatGPT Pro users who pay $200 per month. For this group, the question is whether GPT-4.5 offers sufficient advantages over cheaper models like GPT-4o. OpenAI has indicated that GPT-4.5 will eventually also be made available to users of the ChatGPT Plus plan.

Conclusion

GPT-4.5 brings clear improvements in interaction, creativity, and error reduction, but it is not a fundamental breakthrough. The model is expensive, does not always perform better than alternatives, and raises questions about the scalability of traditional AI training techniques.