OpenAI GPT-4.5 Research Preview

OpenAI has released a research preview of GPT-4.5, its largest and most knowledgeable model to date. This model is designed as a general-purpose system, prioritizing broad knowledge and natural interaction over the specialized STEM-focused reasoning of other models.

Model Architecture and Training

GPT-4.5 builds upon the foundation of GPT-4o, scaling pre-training further to increase its knowledge base. The training process utilized a combination of new supervision techniques alongside traditional methods, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF).

Key Capabilities and Improvements

GPT-4.5 demonstrates improved performance in general-purpose tasks such as writing, programming, and practical problem-solving. Key enhancements include:

• Reduced Hallucinations: The model exhibits fewer hallucinations during task execution.
• Enhanced Emotional Intelligence: GPT-4.5 is designed to feel more natural in its interactions.
• Stronger Alignment: The model shows better alignment with user intent, allowing for more accurate responses to complex prompts.
• Broader Knowledge Base: Increased scaling of pre-training has resulted in a larger repository of internal knowledge.

Safety and Evaluation

OpenAI conducted extensive safety evaluations prior to deployment, following the company's safety process and Preparedness Framework. These evaluations did not reveal any significant increase in safety risks compared to existing models.

Research Preview Status

GPT-4.5 is currently available as a research preview. OpenAI is releasing the model in this stage to identify its strengths and limitations and to observe how users apply the model in unexpected ways.