Last week Openai has officially presented GPT-4.5its most advanced artificial intelligence model to date, known internally with the code name “Orion”capable according to the company to provide a more natural interaction capable of understanding user prompts in a “more human” way. This new version was trained with a quantity of data and computational power never seen before in the history of the company directed by Sam Altmanwith the aim of improving the understanding of language, reduce errors and offer more precise answers. Initial access is reserved for preview of chatgpt pro users (at a cost of about $ 200 per month) and developers who use the paid bees, while chatgpt plus users and teams will be able to experiment with from the next few days.
What makes GPT-4.5 special? His training is based onlearning not superviseda technique that allows him to identify patterns and connections in the data without the developers provide them with explicit instructions. This approach has already shown that it is effective in previous models and has now been brought to a higher level. It must be said, however, that Openai itself admits that the improvements deriving from the increase in the computational scale are starting to slow down and that, despite being very powerful, GPT-4.5 does not exceed the best artificial intelligence models specialized in logical and mathematical reasoning, such as those developed by Deepseek.
The characteristics of GPT-4.5
GPT-4.5 is set in a context in which artificial intelligence is becoming increasingly central to many sectors, from programming to education, passing through support for writing and creativity. The model was developed using a methodology already tested by Openai, i.e. a Large scale training based on non -supervised learning. This technique allows the machine to acquire knowledge without the need for explicit guide by human beings, improving the ability to process information and formulate coherent answers.
On a technical level, the new model stands out for a wider and deeper knowledge in the world, which should guarantee more accurate answers and a lower tendency to produce incorrect errors or information, a phenomenon known as “hallucinations of AI”, as shown by some tests conducted internally by the company. Openii also worked for Increase the emotional intelligence of GPT-4.5refining his ability to understand the shades of language and users’ intentions. This translates into more fluid interactions and more natural conversationsmaking it particularly useful in assistance, programming and resolution of complex problems.
Sam AltmanCEO of Openai, in describing the new model said:
It is the first model that gives me the feeling of talking to a caring person. There were several moments when I sat on the chair and I was amazed in receiving really good advice from an AI.
The critical points of GPT-4.5
Despite the interesting peculiarities of his new model, Openai said clearly that GPT-4.5 is not the most advanced model in all fields. In particular, in the tests that evaluate the ability of logical and mathematical reasoning, it has not reached the levels of some competitors specialized in these areas. Models such as those developed by Deepseek-R1 passing through O1 and O3-Mini of the same Openai, these have achieved better results in specific benchmark. Even in the field of coding, GPT-4.5 proved to be effective, but has not exceeded the most advanced models dedicated to programming.
A critical aspect of GPT-4.5 is his high cost. For developers who wish to access the API, Openai has set a price of 75 dollars For every million input token e 150 dollars For every million Output token. These are decidedly higher costs compared to GPT-4O, which has a price of 2.50 dollars per million input token e 10 dollars per million output token. This could make GPT-4.5 less accessible to many independent companies and developers, limiting their adoption on a large scale.
Despite the costs, Openai has decided to release GPT-4.5 as a research preview, to allow users to test their skills and identify potential applications. The company is aware that the future of artificial intelligence is not only based on the increase in computational power, but also on new approaches to learning and reasoning. For this reason, Openii is exploring different strategies to further improve its technologies, combining learning not supervised with models specialized in structured reasoning.
As for the model safetyOpenai has implemented new supervision techniques to reduce the risk of improper uses and improve the alignment of the model to human intentions. This is particularly important given the growing use of AI in sensitive contexts, such as research support, writing professional documents and assistance to business decisions.
If you want to deepen the peculiarities of the model, we leave you to the video (in English) relating to its official presentation made available by Openai.
https://www.youtube.com/watch?v=cfryp0nitz8
The availability of GPT-4.5
On a practical level, users of Chatgpt Procan already access GPT-4.5, selecting it directly in the interface of the service, thanks to the appropriate selector of the models in the chatgpt. Starting this week, also users of Chatgpt Plus And Team they should be able to try it, while subscribers to Chatgpt Enterprise and EDU they will have to wait a little longer. The model offers advanced features, including the possibility of analyzing documents, images and writing code, but at the moment does not support the realistic vocal mode that is present in GPT-4O. For developersGPT-4.5 is already available through API.
