GPT-4.5 released: More human and less hallucinatory
GPT-4.5 is not a Frontier model, they say. However, users should notice a clear difference in its responses compared to previous models.
(Image: OpenAI)
OpenAI has published a research preview of its AI model GPT-4.5. The new AI model is more of a fine-tuning than a big hit. Nevertheless, it shows where the journey of OpenAI's Large Language Models (LLMs) is headed. And so users are already getting a small foretaste of GPT-5, which is rumored to be released in May.
GPT-4.5 will be closer to the user in various ways. A higher level of emotional intelligence, or EQ for short, should help the model to better understand nuances, subtle hints and implicit expectations in communication. The model itself should come across as more human – and get to the point more quickly in its responses. In its press release, OpenAI uses examples to show how the new model responds briefly and succinctly compared to GPT-4o, while the older model gets bogged down in detailed explanations.
Fewer hallucinations
According to OpenAI's description, GPT-4.5 should be particularly useful for creative writing, but also for programming and practical problem solving. Compared to GPT-4, the computing efficiency has been improved 10-fold.
The new model should also hallucinate less. By publishing it as a research preview, OpenAI wants to find out for itself where the strengths and weaknesses of the model lie. With the scaling of unsupervised learning, GPT-4.5 should improve its ability to recognize patterns, draw connections and generate creative insights, according to OpenAI. Unlike reasoning models, models like GPT-4.5 should have a deeper intuitive understanding of the world.
Also weaker in some areas
Compared to o1 and o3-mini, GPT-4.5 is said to be weaker in some areas and is therefore not considered a "frontier model" by OpenAI. It also does not currently support multimodal functions such as voice mode, video or screen sharing.
Videos by heise
GPT-4.5 will initially be made available to ChatGPT Pro users. Plus and Team users will follow next week, followed by Enterprise and Edu customers. The model will also be accessible via the Azure AI Foundry platform and is available for developers in the Chat Completions API, Assistants API and Batch API.
(mki)