GPT-4.5 scales unsupervised learning for more natural conversation

OpenAI introduced GPT-4.5 today as a research preview for ChatGPT Pro users and developers. The company describes it as its largest and most knowledgeable GPT model, built by scaling unsupervised learning rather than the deliberate reasoning approach used in its o-series.

The release supports file and image input, web search and a programming interface, while voice and video are absent at launch. OpenAI reports lower hallucination rates on its SimpleQA evaluation and stronger human preference in writing, practical knowledge and professional queries.

GPT-4.5 is computationally expensive, and OpenAI says it is evaluating whether to keep serving it through the API. That caveat makes the preview an economic experiment as well as a technical one. A model that feels more natural may be useful for writing and collaboration, but fluency still cannot substitute for verified facts or task-specific evaluation.