On Thursday, OpenAI rolled out the release of GPT-4.5, the newest version of its cornerstone language model. This updated edition, a major step up in the world of generative AI, offers improved pattern recognition, enhanced natural language interfaces, and reduced “hallucinations”—mistakes or inventions that have bedeviled previous models.
GPT-4.5 is OpenAI’s biggest model to date, although the firm won’t reveal its size or the amount of computing power needed to train it. Available to ChatGPT Pro customers for $200 a month, and eventually to other paying users, the model differs from the earlier versions in terms of design philosophy. OpenAI noted that GPT-4.5 does not try to annihilate current benchmarks but presents a more “reflective” and natural conversation experience.
Sam Altman, CEO of OpenAI, described GPT-4.5 as the first model that is like communicating with a “thoughtful person.” Though not a reasoning-based model, it presents a new type of intelligence that Altman feels has a special “magic.” The new version has real-time search, image uploads, and file sharing but currently omits voice mode, video, and screen sharing.
The main upgrade to GPT-4.5 is its basis on scaling the pretraining stage, an unsupervised learning method that OpenAI has developed over the long term to optimize its models. The shift will lead to a wider knowledge base of the world, fewer mistakes, and higher reliability in consumer interactions. Nevertheless, OpenAI cautioned that GPT-4.5 costs more computationally to serve up than its older versions, rendering it a high-end product to pay for special features.
The GPT-4.5 release represents a critical point in OpenAI’s development plan. The company claims it will be the final “non-chain-of-thought” model. The future of AI development will increasingly rely on reasoning ability to enable models to provide more considered, precise responses based on the context of a question.
OpenAI has stated that although GPT-4.5 remains in a research preview state, it will continue to improve the model’s reasoning capabilities, with future versions such as GPT-5 incorporating even more sophisticated features.
