OpenAI promises GPT-4.5 will "hallucinate less"

The hotly anticipated GPT-4.5 model has finally arrived as a research preview and will soon power OpenAI’s ChatGPT for everyone.

The company describes it as its “most knowledgeable model yet,” with improvements in writing, world knowledge, and pattern recognition. OpenAI says conversations with GPT-4.5 will feel more natural, and the model is better at programming and problem-solving.

According to a leaked document, GPT-4.5 is more computationally efficient than GPT-4 but does not outperform OpenAI’s o1 or o3-mini models in certain evaluations. The company has since removed this note from an official update.

Sam Altman, OpenAI’s boss, has previously hyped the model as the “closest thing” we have to artificial general intelligence (AGI). It was trained using synthetic data on Microsoft Azure AI supercomputers and new supervision techniques to “hallucinate less” compared to GPT-4o.

Following its release for Pro users, GPT-4.5 will expand to Plus and Team users next week, with Enterprise and Edu users gaining access later. It is also available on Microsoft’s Azure AI Foundry alongside models from Stability AI and Cohere.

Altman says that GPT-4.5 “isn’t a reasoning model and won’t crush benchmarks” but rather “a different kind of intelligence and there’s a magic to it.”

The Microsoft-backed company also plans to launch GPT-5 by late May, as The Verge has exclusively revealed.

As expected, OpenAI is discontinuing the o3 model and integrating its functionality into GPT-5 as part of a broader plan to remove the model picker and create a unified AI system.

The transition will start with GPT-4.5, which will pave the way for GPT-5 to combine capabilities from both the GPT-series and o-series models.