OpenAI has released GPT-4.5 as a research preview, describing it as the company's largest and best chat model to date. The model is initially available to ChatGPT Pro users and developers, with access expanding to Plus and Team users starting next week.

  • GPT-4.5 (also known as Orion) uses traditional pre-training scaling rather than the reasoning approach of the o1 series.
  • It costs $75 per million input tokens and $150 per million output tokens, significantly more than GPT-4o or o1.
  • In benchmarks, it achieves 62.5% accuracy on SimpleQA and reduces hallucination rates to 37.1%, outperforming predecessors in some areas but lagging behind reasoning models like o3-mini in STEM tasks.
  • Human evaluations show testers preferred GPT-4.5 over GPT-4o across creative, professional, and everyday categories.

OpenAI views GPT-4.5 as a foundation for future intelligence, combining pre-training scaling with reasoning capabilities expected in GPT-5.