OpenAI has unveiled GPT-4.5, their latest and most extensive language model, launching it initially as a research preview for ChatGPT Pro subscribers. While heralded as their “most knowledgeable model to date,” OpenAI cautions that GPT-4.5 isn’t a revolutionary “frontier” model and might not surpass the performance of their specialized reasoning models like o1 or o3-mini in certain evaluations.
GPT-4.5 brings enhancements to writing capabilities, a broader understanding of world knowledge, and a more polished conversational “personality” compared to its predecessors. OpenAI emphasizes that interactions with GPT-4.5 will feel more “natural,” highlighting its improved ability to discern patterns and forge connections, making it well-suited for tasks like writing, coding, and practical problem-solving.
However, OpenAI clarifies that GPT-4.5 doesn’t introduce a sufficient number of entirely novel capabilities to qualify as a “frontier” model. The model’s primary focus is on boosting computational efficiency, reportedly by more than tenfold compared to GPT-4, and reducing hallucinations. While trained using new supervision techniques alongside established methods like supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to GPT-4o, it was also trained using data from their o1 reasoning model, codenamed Strawberry.
A key advantage of GPT-4.5 is its significantly reduced tendency to “hallucinate,” or generate inaccurate information, surpassing GPT-4o and approaching the reliability of their o1 model. OpenAI has also prioritized aligning GPT-4.5 to be a more effective collaborator, aiming for conversations that feel warmer, more intuitive, and emotionally nuanced. Human testers have consistently rated GPT-4.5 as superior to GPT-4o across various evaluation categories.
OpenAI CEO Sam Altman acknowledges that GPT-4.5 is a substantial and resource-intensive model, emphasizing that its primary strength lies in its refined performance rather than achieving benchmark-breaking scores. Following its initial launch for Pro users, GPT-4.5 will be rolled out to Plus, Team, Enterprise, and Edu users in subsequent phases. It is also immediately available on Microsoft’s Azure AI Foundry platform, alongside newly released models from Stability, Cohere, and Microsoft.
OpenAI highlights that GPT-4.5 produces more accurate responses and exhibits fewer hallucinations compared to their previous models. It was previously reported that OpenAI was planning to launch GPT-4.5 by the end of February, with GPT-5 expected as soon as late May. OpenAI CEO Sam Altman has described GPT-5 as a “system that integrates a lot of our technology,” which will include their new o3 reasoning model, teased during their December announcements. While OpenAI released o3-mini last month, the full o3 model will be integrated into the upcoming GPT-5 system. This strategy aligns with OpenAI’s long-term goal of combining their various large language models to create a more capable system, potentially leading to artificial general intelligence (AGI).