OpenAI has announced the release of a preliminary version of GPT-4.5, a new evolution of its GPT model designed for researchers, developers, and Pro users worldwide. This iteration of the GPT model focuses on improving unsupervised pre-training, with the primary goal of enhancing response reliability and refining the understanding of user intentions and emotions.
According to OpenAI, GPT-4.5 offers notable improvements in its ability to identify complex patterns, establish connections, and generate creative ideas, while reducing the hallucinations commonly observed in previous models. Preliminary tests indicate that interactions with GPT-4.5 are perceived as more natural by users.
These improvements primarily result from an increase in the scale of unsupervised pre-training, made possible by Microsoft's Azure AI supercomputers, as well as architectural optimizations. The model notably displays increased accuracy in factual responses (62.5% on the SimpleQA index compared to 38.2% for GPT-4o), and a reduced hallucination rate of 37.1%.
GPT-4.5 also integrates better consideration of emotional context, demonstrating greater emotional intelligence (EQ) for better understanding of implicit nuances. It is particularly effective in applications such as creative writing, programming, or solving practical daily problems.
However, OpenAI specifies that GPT-4.5 remains a research preview whose purpose is precisely to evaluate its capabilities and limitations in various contexts before a potential wider release.
GPT-4.5 is now available for ChatGPT Pro users and will gradually become accessible to Plus and Teams users starting next week, then to Enterprise and Edu users later. Developers already have access to the model via OpenAI's Chat Completions, Assistants, and Batch APIs.
Watch out for the cost!
GPT-4.5 does not replace GPT-4o, particularly due to higher operational costs, but it is intended for specific use cases requiring high contextual sensitivity and better adaptation to human interactions.
Inputs cost $75/1M Tokens (compared to $2.5/1M Tokens for GPT-4o)
Outputs cost $150/1M Tokens (compared to $10/1M Tokens for GPT-4o)