OpenAI has officially launched GPT-4.5, code-named Orion, which has become its heaviest AI model to date. As the company builds it with increased computing power and vast amounts of data, this has been perceived as a considerable leap forward for the company. But in an unexpected turn, OpenAI eliminated a line from the original white paper that stated GPT-4.5 was not a frontier model. Subscribers to ChatGPT Pro and developers on paid API plans accessed it this Thursday, and it will become available for everyone over the coming weeks.
Although GPT-4.5 has displayed remarkable advancements in some areas, such as factual accuracy and social appropriateness, OpenAI admits to the model’s drawbacks in these aspects, particularly when compared to newer reasoning models from companies such as DeepSeek and Anthropic.
Building on earlier work like the radically improved unsupervised learning practices used in both GPT-3 and GP-4, it is probably going to continue being part of OpenAI’s effort to enhance the performance of AI models with respect to data-input weight. It’s still highly imperfect, though colossal: OpenAI itself notes that some areas are beginning to plateau in performance as scaling data and computing power produce no longer meaningful increases in all domains.
It certainly excels on benchmarks like SimpleQA but lags behind doing complex coding and scholastic problems compared to reasoning-centered models. Another thing working against GPT-4.5 is its computational cost, as it affects how OpenAI will have to juggle between making great strides in performance while not trespassing on sustainability in the long term with its API offerings.
Not that it doesn’t have those drawbacks, but certainly, it has the most promising results when it comes to creative exercises such as writing and design as well as showing emotion and social appropriateness more than any other model. OpenAI is still looking into what it can do and where it falls short, for the benchmarks are not considered a very good indicator of how it might fare in a more realistic environment.
It may not capture the high ground yet, but GPT-4.5 is already billed as the first step into the future, as it apparently sets the pace for more powerful models such as GPT-5, which may feature reasoning-based enhancements. OpenAI’s exploration into combining its GPT models with reasoning AI could be a major variable in solving pre-training scaling laws, as the industry will speed toward more sophisticated and consistent AI systems.
