OpenAI released GPT-4.5, its newest large language model (LLM), as a research preview for ChatGPT Pro users. Described as their 'most knowledgeable model yet', OpenAI clarifies that it's not a 'frontier' model and may not outperform previous models like o1 or o3-mini in some areas. The model boasts enhanced writing abilities, improved world knowledge, and a more 'natural' interaction.
GPT-4.5 shows improvements in pattern recognition and connection-drawing, making it suitable for writing, programming, and problem-solving. However, OpenAI initially stated that it lacked significant 'frontier' capabilities and was underperforming compared to earlier models on certain benchmarks; this statement was later removed from an updated document.
The model was reportedly trained using a combination of new supervision techniques and traditional methods like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF).
GPT-4.5 will be rolled out to Plus, Team, Enterprise, and Edu users in phases, and is already available on Microsoft's Azure AI Foundry. OpenAI's CEO, Sam Altman, acknowledged the model's high cost and confirmed its limitations in benchmark tests. The release precedes the anticipated launch of GPT-5, which is expected to integrate OpenAI's new o3 reasoning model and aim towards artificial general intelligence (AGI).