GPT-4.5 is here, marking a significant advancement in the world of generative AI. OpenAI’s newest model is not only larger and more compute-intensive, but it also promises to enhance the understanding of user prompts in ChatGPT. For those eager to be among the first to experience this innovative technology, a subscription to OpenAI’s $200-per-month ChatGPT Pro service is required, as the model is currently available as part of a research preview.
The launch of GPT-4.5 comes amidst a flurry of new AI model releases in 2025. Notably, Anthropic has introduced a hybrid reasoning model for its Claude chatbot, while Chinese researchers at DeepSeek have made waves in Silicon Valley with a powerful model developed on a surprisingly low budget. In response, OpenAI released a “mini” version of its reasoning model just a month prior. As these developments unfold, OpenAI has committed to investing billions into building the AI infrastructure necessary to support the development of even more advanced models.
The launch of GPT-4.5 reaffirms OpenAI’s strategy of scaling its models for better performance. This approach contrasts sharply with other recent innovations, such as DeepSeek’s R1, which aimed to achieve high performance with minimal resources. OpenAI believes that a maximalist approach to model development can better capture the nuances of human emotions and interactions. Researchers have suggested that the larger size of GPT-4.5 may also lead to a reduction in the frequency of hallucinations compared to previous versions. Mia Glaese, who leads OpenAI’s alignment and human data teams, stated, “If you know more things, you don't need to make things up.”
While specific details regarding the size and compute intensity of GPT-4.5 have not been disclosed by OpenAI, Pro users are currently experiencing the new model, with rollouts planned for Plus and Team users next week, and for Enterprise and Edu users the week after. This version supports a range of features including web search, canvas usage, and file and image uploads, though it is not yet compatible with the AI Voice Mode.
In its announcement, OpenAI shared academic benchmark results indicating that GPT-4.5 has been outperformed by the o3-mini model in math and slightly lagged in science. However, it did outperform its predecessor on language benchmarks. Researchers emphasize that these metrics do not tell the whole story, with Glaese noting, “We would expect the difference in 4.5 to be similar to the experience difference of 4 to 3.5.” Users engaging with topics such as writing or programming may find that GPT-4.5 provides enhanced results, with a more natural flow in back-and-forth interactions.
Unlike models released under OpenAI’s “o” series, GPT-4.5 is not classified as a reasoning model. OpenAI CEO Sam Altman mentioned in February that GPT-4.5, internally referred to as Orion, would be the last non-chain-of-thought model. Nick Ryder, who oversees the foundations-in-research team, clarified that this statement reflects a shift towards streamlining OpenAI’s product roadmap rather than its research roadmap. Future releases aim to provide users with a more integrated experience, allowing the AI to determine the best model for responding to user prompts.
As OpenAI navigates increasing competition in the AI landscape, it aims to maintain its position at the forefront of technological innovation. To achieve this, the company is focusing on enhancing pretraining processes. “By increasing the amount of compute we use, by increasing the amount of data we use, and focusing on really efficient training methods,” Ryder explained, “we push the frontier of unsupervised learning.” With the substantial size of GPT-4.5, questions have arisen about whether this will complicate the understanding of the model's internal processes. However, Ryder believes that existing methods for interpreting smaller models will apply equally to larger models.
As part of WIRED’s ongoing coverage of new software releases, I will be testing GPT-4.5 to evaluate its performance against its competitors and predecessor models. Due to OpenAI’s emphasis on the model’s potential strengths—such as enhanced intuition, emotional intelligence, and aesthetic taste—direct comparisons may prove challenging. Nonetheless, OpenAI is committed to developing AI that can match the productivity of human workers while excelling in essential soft skills.
