Hands-On With GPT-4.5, OpenAI’s Most Powerful Model Yet

While the improvements feel as incremental as its name suggests, GPT-4.5 is still OpenAI’s most ambitious drop to date. Released in late February as a research preview—which essentially means OpenAI sees this as a beta version—GPT-4.5 uses more computing power than its previous models and was trained on more data.

So, just how big is the GPT-4.5 research preview? Who knows—since the developers won’t say. And where did this additional training data come from? Their lips are zipped on that as well. To borrow a line from Apple TV’s hit show Severance, right now OpenAI is positioning the alleged improvements in this new model as mysterious and important.

When comparing AI benchmark tests from competitors’ models as well as OpenAI’s “reasoning” releases, the benefits of using GPT-4.5 are not immediately clear. Though, in the model’s system card and in a previous interview with WIRED, the OpenAI researchers who worked on GPT-4.5 claimed improvements can be felt in the anthropomorphic aspects of the model, like a stronger intuition and a deeper understanding of emotion. After sitting in OpenAI’s office last year and listening to leadership talk about the startup’s plan to further productize ChatGPT as useful software, this was not the release I expected in 2025. Rather than take a more utilitarian approach, this model attempts to be more emotional.

OpenAI has steadily grown its number of enterprise contracts, so the company could be expected to put out tentpole releases with baked-in, practical applications, especially inside the most expensive and powerful version of its chatbot. However, GPT-4.5 is more aligned with the output from an academic research group pouring everything they have into chasing artificial general intelligence, a theoretical version of the algorithm that’s deft enough to replace white-collar workers and practically God-like in its ability to process information.

While OpenAI would argue that these two paths are intertwined and equally important, if your short-term goal is to make money from ChatGPT, last week’s belabored release makes no sense; it’s super expensive and offers marginal gains only seasoned chatbot users may notice. But if your overarching mission is to build beneficial AGI, which is still OpenAI’s core objective, then mimicking the nuances of human emotions and soft skills remains a critical area for improvement. It’s where the company could hold onto its leading position as additional competitors in the generative AI race, like the vastly cheaper R1 model from DeepSeek, push forward on other innovations.

How to Access GPT-4.5

As with most of the new features and models that arrive for ChatGPT, OpenAI’s paid subscribers will be the first to gain access to GPT-4.5. In this case, OpenAI is unlocking access first for ChatGPT Pro subscribers who pay a hefty $200 a month. The large rollout of GPT-4.5 to the other paid tiers—Plus, Team, Enterprise, and Edu—will happen during this week and the next. Past models have eventually trickled down to the free version of ChatGPT as well, but the company does not yet have a plan to release GPT-4.5 to all users, due to its size and computing requirements.

When it becomes available in your account, GPT-4.5 will be one of the many options nestled in the model dropdown menu that appears when you click the word ChatGPT at the top of the screen. In my Pro account, this raised the current total number of available models to a whopping nine different choices that I now have to pick between. OpenAI developers have told me that they hope to significantly streamline that process in the future and have the AI tool pick which model is best suited for each prompt the user types or speaks.

First Impressions of OpenAI’s New Model

The draft headline I put on this article was “With GPT-4.5, OpenAI Gets Lost in the AGI Sauce.” And while no headline featuring the model’s name is going to feel poetic, that’s a bit of mess. Writing strong, succinct headlines is a difficult skill requiring clear communication as well as a level of aesthetic taste—often involving the input of multiple editors before the perfect message is conveyed. I was curious about whether ChatGPT would be able to punch up that headline, so I tried to do that using both the newest model and GPT-4o, a past release the company describes as “great for most tasks.”

Among all the intangible improvements, GPT-4.5 was much more capable at writing a compelling headline. GPT-4o’s outputs were less interesting and had less variety overall, with the exception of this nonsensical banger: “With GPT-4.5, OpenAI Keeps One Foot in the Future, One in the Chatbot.” Here is a much better punch-up provided by the new model: “With GPT-4.5, OpenAI Trips Over Its Own AGI Ambitions.” It’s fairly similar to the original, but potentially more clear for readers. After some consternation, the human editors at WIRED decided to go with the headline generated by GPT-4.5. Fair enough.

Switching gears, I asked why the price of a dozen eggs is rising even higher during the beginning of Trump’s presidency than it was under Biden, primarily to see which model would be more successful at analyzing web articles about a political topic. The differences here were more subtle, but GPT-4o seemed prone to lecture me and repeat itself, whereas GPT-4.5 did a better job of understanding my intent and succinctly representing multiple viewpoints.