OpenAI has recently unveiled GPT-4.5, its latest and most expansive AI language model to date. Initially, this model will be accessible as a research preview exclusively for ChatGPT Pro users. OpenAI has touted GPT-4.5 as its “most knowledgeable model yet,” but has also cautioned that it does not qualify as a frontier model and may not perform as effectively as its predecessors, specifically the o1 and o3-mini models.
The enhancements in GPT-4.5 include superior writing capabilities, an enriched understanding of world knowledge, and what OpenAI describes as a “refined personality” compared to earlier iterations. The organization claims that interactions with GPT-4.5 will feel more “natural,” highlighting the model’s improved ability to recognize patterns and establish connections. This makes GPT-4.5 particularly well-suited for tasks such as writing, programming, and addressing practical problems.
Despite these advancements, OpenAI has clarified that GPT-4.5 does not introduce sufficient new capabilities to be classified as a frontier model. According to a leaked document prior to the official announcement, OpenAI stated, “GPT-4.5 is not a frontier model, but it is OpenAI’s largest LLM, improving on GPT-4’s computational efficiency by more than 10 times.” The document also noted that GPT-4.5 does not present seven new frontier capabilities compared to previous reasoning releases and that its performance falls short of the o1, o3-mini, and deep research models in most preparedness evaluations.
This particular mention has since been removed from an updated version of the document. It was previously reported that OpenAI utilized its o1 reasoning model, codenamed Strawberry, to train GPT-4.5 using synthetic data. The training process for GPT-4.5 involved new supervision techniques alongside traditional methods such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to those employed for GPT-4o.
Notably, GPT-4.5 exhibits a significant reduction in hallucinations compared to GPT-4o and shows slightly improved performance over the o1 model. Raphael Gontijo Lopes, a researcher at OpenAI, emphasized during a livestream that “we aligned GPT-4.5 to be a better collaborator, making conversations feel warmer, more intuitive, and emotionally nuanced.” Human testers evaluated GPT-4.5 against GPT-4o, and the results indicated that GPT-4.5 outperformed its predecessor in nearly every category.
In a post on X, OpenAI CEO Sam Altman acknowledged the scale and cost of GPT-4.5, stating that it is a “giant, expensive model” that “won’t crush benchmarks.”
Following its initial rollout for Pro users, OpenAI plans to extend access to Plus and Team users in the coming week, with subsequent availability for Enterprise and Edu users. Additionally, GPT-4.5 is now accessible on Microsoft’s Azure AI Foundry platform, alongside new models from Stability, Cohere, and Microsoft.
OpenAI had originally aimed to launch GPT-4.5 by the end of February, with plans for GPT-5 to follow as early as late May. Altman has described GPT-5 as a “system that integrates a lot of our technology,” which will incorporate OpenAI’s new o3 reasoning model, a feature teased during the company’s 12 days of Christmas announcements in December.
While OpenAI released the o3-mini model last month, the full o3 model will only be included in the upcoming GPT-5 system. This strategy aligns with OpenAI’s overarching goal of merging its large language models to ultimately develop a more capable model that could be recognized as artificial general intelligence (AGI).
