On February 27, 2025, OpenAI officially released its latest language model GPT-4.5, which was the first to be open to ChatGPT Pro users as a research preview version, with a subscription fee of US$200 per month. Built on GPT-4o, this model is widely regarded as the largest and most knowledgeable language model to date. GPT-4.5 has achieved significant improvements in its inference ability, emotional intelligence, practical functions such as writing and programming, marking another major breakthrough in artificial intelligence technology. OpenAI said that in the next few weeks, GPT-4.5 will be gradually opened to Plus, Team, Enterprise and Edu users, further expanding its application scope.
GPT-4.5 has shown remarkable progress in many areas. According to the system card released by OpenAI, the model significantly reduces the “illusion” phenomenon (i.e., error or fictional information) through supervised fine-tuning (SFT) and reinforcement learning (RLHF) training, and improves the naturalness of emotional intelligence and user interaction. For example, in PersonQA test, the hallucination rate of GPT-4.5 was only 19%, which is much lower than the 52% of GPT-4o. In terms of user experience, GPT-4.5 performs well in writing and programming tasks, producing more creative and aesthetically intuitive content while understanding user intentions more accurately. X user Simon Willison said after the trial: "It performs better in terms of depth of knowledge and content generation (such as drawing the Pelican), and has a more natural interactive experience." In addition, the model supports file and image uploads and integrates ChatGPT's canvas tool, but currently does not support multimodal features such as voice, video or screen sharing.
OpenAI revealed that GPT-4.5's training data comes from diverse channels, including public data, proprietary partner data, and internal customized data sets, which are strictly filtered to ensure quality and protect privacy. In terms of security, the model is equipped with a Moderation API and a classifier to effectively prevent harmful content from being generated. In third-party assessments, Apollo Research and METR tested the risk of fraud and autonomy of GPT-4.5, and the results showed that its overall risk was "medium". Especially in the fields of chemistry, biology, radioactive nuclear materials (CBRN) and persuasion, the risks are moderate, while cybersecurity and model autonomy are low. OpenAI has taken steps such as pre-training to filter CBRN data and monitor potential abuse to ensure security.
Feedback from the first batch of users showed that GPT-4.5 was well received in terms of knowledge breadth and creativity. X user @pckraftwerk praised its "significantly enhanced writing and coding capabilities, with reduced hallucination phenomena". However, the speed issue has also attracted attention. Simon Willison pointed out that when used through the API, GPT-4.5 is slow to respond, which may affect real-time application scenarios. He also revealed that the model uses 10 times more training computing resources than GPT-4, and the performance improvement is huge investment behind it. In addition, OpenAI made it clear that GPT-4.5 is not designed for deep reasoning, and it is recommended that users choose the o1 or o3-mini model when complex reasoning is needed. This positioning has also triggered some discussion.
As OpenAI's latest masterpiece, GPT-4.5 is believed to promote productivity innovation in areas such as writing, programming and problem solving. Its multilingual performance is also impressive. In the MMLU benchmark test, the English accuracy rate reached 0.896, Spanish 0.8840, and Arabic 0.8598, surpassing GPT-4o in all aspects. Industry insiders predict that with the gradual promotion of GPT-4.5, its powerful functions and high security will further consolidate OpenAI's leading position in the AI field. In the future, OpenAI plans to optimize the model based on user feedback, and may expand functions such as voice and video to further enrich its application scenarios.
The release of GPT-4.5 not only demonstrates OpenAI's technical ambitions, but also highlights its emphasis on security and user experience. Although the speed problem remains to be solved, this model undoubtedly opens a new chapter in the development of AI. As OpenAI said in its statement: "We hope that GPT-4.5 will bring users a smarter and more reliable assistant experience."