OpenAI has released a research preview of GPT-4.5, describing it as its largest and most knowledgeable model to date. Built on the GPT-4o architecture, the new model scales pre-training further and is designed to be more general-purpose than previous STEM-focused reasoning models.
The system card outlines several key aspects of the release:
- Training utilizes new supervision techniques combined with supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF).
- Early testing indicates more natural interaction, stronger alignment with user intent, improved emotional intelligence, and fewer hallucinations.
- Safety evaluations found no significant increase in risk compared to existing models across categories like CBRN, cybersecurity, persuasion, and model autonomy.
OpenAI is sharing this preview to better understand the model's strengths and limitations while exploring how users might utilize its capabilities.