OpenAI has released GPT-4.5, its newest and biggest AI language model. It is available first as a research preview for ChatGPT Pro users. OpenAI says GPT-4.5 is its “most knowledgeable model yet,” but also warns that it is not a major breakthrough in AI and may not perform as well as other models like o1 or o3-mini.
GPT-4.5 has better writing skills, improved knowledge, and a more refined personality. OpenAI says it will feel more “natural” to use and is better at recognizing patterns, making connections, and solving problems. This makes it useful for writing, coding, and everyday problem-solving.
However, OpenAI says GPT-4.5 does not bring enough new features to be considered a top-tier AI model. A leaked document stated that GPT-4.5 improves GPT-4’s efficiency by 10 times but does not introduce seven major new abilities. OpenAI later removed this information from its updated document. It was also reported that OpenAI used synthetic data from its o1 model, called “Strawberry,” to help train GPT-4.5. The company says it used a mix of new supervision techniques, fine-tuning, and reinforcement learning from human feedback (RLHF) to train the model.
Despite its limits, GPT-4.5 makes fewer mistakes than GPT-4o and slightly fewer than o1. OpenAI researcher Raphael Gontijo Lopes said the model is more natural in conversations and feels “warmer and more intuitive.” Human testers also rated GPT-4.5 higher than GPT-4o in almost all areas.
OpenAI CEO Sam Altman noted that GPT-4.5 is a “giant, expensive model” but admitted that it “won’t set new records.” GPT-4.5 will first be available to Pro users, then to Plus and Team users next week, followed by Enterprise and Edu users. It is also available in Microsoft’s Azure AI Foundry.
Reports suggest OpenAI plans to launch GPT-5 as early as May, integrating the o3 reasoning model, which was teased last December. The company aims to combine its AI models over time to create an advanced system that could lead to artificial general intelligence (AGI).
Source: theverge
