The last few days have been very busy with new model releases, and OpenAI is closing the week with a research preview of GPT-4.5. The company is positioning the new system as its biggest and best model for chat. During early testing, according to OpenAI, people found GPT-4.5 to be a more natural conversationalist, able to convey warmth and demonstrate a kind of emotional intelligence.
In one example cited by OpenAI, a person tells ChatGPT that they are going through a difficult time after failing a test. While the company’s previous models, including the GPT-4o and o3-mini, might have sympathized with the person before offering a long list of unsolicited advice, the GPT-4.5 takes a different tact. “Do you want to talk about what happened, or do you just need a distraction? I’m here either way,” the chatbot says while running on GPT-4.5.
The improvements shown by GPT-4.5 are the result of OpenAI’s advances in unsupervised learning. In unsupervised learning, a machine learning algorithm is given an unlabeled dataset and is left to find patterns and insights on its own. GPT-4.5 doesn’t “think” like the company’s most advanced reasoning models, but while training the new model, OpenAI made architectural improvements and gave it access to more data and computing power. “The result is a model that has broader knowledge and a deeper understanding of the world, resulting in fewer hallucinations,” the company says.
Speaking of reducing hallucinations, OpenAI measured how GPT-4.5 is better in this regard. When run through SimpleQA, a benchmark developed by OpenAI that tests large language models for their ability to answer “simple but complex questions,” GPT-4.5 outperformed o3-mini, GPT-4o, and even o1 with a 37.1 percent hallucination rate. Obviously, the new model does not completely solve the problem of AI hallucinations, but it is a step in the right direction.
Despite its relative advantages over GPT-4o and o3-mini, GPT-4.5 is not a direct replacement for these models. Compared to OpenAI reasoning systems, GPT-4.5 is a “more versatile, innately smarter model”. In addition, it is not multimodal like GPT-4o, which means that it does not work with features such as voice, video, or screen sharing. In addition, it is “a very large and resource-intensive model.”
It’s best to think of GPT-4.5 as a stepping stone to the systems OpenAI plans to offer in the future. In fact, Sam Altman said as much earlier this month when he shared the company’s roadmap, noting that GPT-4.5 will be “our last model without a thought chain” – referring to the fact that the new system doesn’t solve problems by tackling them step by step, as OpenAI’s reasoning models do. Its successor, GPT-5, is likely to integrate many of OpenAI’s latest technologies, including its advanced o3 model. OpenAI confirmed this today, stating that it plans to bring “the unique strengths of GPT-4.5, including broader knowledge, stronger intuition, and better ‘equalizer’ to all users in future models.”
In the meantime, ChatGPT Pro subscribers can start using GPT-4.5 today, while Pro and Team users will have access starting next week.
