Microsoft-backed OpenAI has launched it latest GPT-4.5 model as a limited preview as it prepares to shift to a new approach with its next iteration, GPT-5, later this year.

The new model is to be released first to a limited number of users in a “research preview” to those paying $200 (£159) a month for ChatGPT Pro.

It said it would incorporate feedback from those users before launching the model to Plus and Team users later this week and then Enterprise and Education users later.

Training techniques

The model is available on Microsoft’s Azure AI Foundry platform, which also offers new models from Stability, Cohere and Microsoft.

The development of GPT-4.5 has been difficult after it faced problems including finding new sources of high-quality training data.

The start-up used a process called post-training in which it incorporated human feedback to improve responses and refine the nuances of how it interacts with users.

It also used its o1 “reasoning” model to train GPT-4.5 with synthetic data.

GPT-4.5 was trained “using new supervision techniques combined with traditional methods like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to those used for GPT-4o,” OpenAI said.

It said GPT-4.5 hallucinates less than GPT-4o and slightly less than the o1 reasoning model.

‘Emotional nuance’

Reasoning models produce responses more slowly and spend more time “considering” what they will say, which is intended to produce more accurate answers and reduce errors such as “hallucination”, or the production of false information.

“We aligned GPT-4.5 to be a better collaborator, making conversations feel warmer, more intuitive, and emotionally nuanced,” said OpenAI researcher Raphael Gontijo Lopes during a streamed launch event.

With GPT-5, OpenAI plans to take a different approach by combining its GPT-series models with its o-series reasoning models, with the ChatGPT chatbot itself deciding which model to use.

ChatGPT itself currently offers users a choice of the model they would like to use, something OpenAI said was overly complex.