Highlights:

  • Compared to its predecessor, GPT-4o, OpenAI’s latest language model demonstrates an improved ability to interpret user input.
  • GPT-4.5 is powered by an AI development technique known as unsupervised learning, which plays a key role in its enhanced capabilities.

OpenAI introduced GPT-4.5, a general-purpose large language model that it describes as its most advanced and capable version to date.

The company provides two distinct collections of large language models. The first collection consists of models specifically designed for reasoning-intensive tasks, such as code generation. The second collection, by contrast, is not optimized for advanced reasoning but instead focuses on text generation tasks that emphasize creativity and general knowledge. GPT-4.5 belongs to this latter category.

Compared to its predecessor, GPT-4o, OpenAI’s latest large language model demonstrates an improved ability to interpret user input. Recently, OpenAI explained that GPT-4.5 can recognize “subtle cues or implicit expectations” within prompts. This enhancement allows the model to better align its responses with user intentions and produce more contextually appropriate outputs.

According to OpenAI, GPT-4.5 also introduces advancements in emotional intelligence over GPT-4o, which enables it to perform certain tasks with greater effectiveness. For example, the model is now more reliable in determining whether a user is expecting a summary or a more detailed response to a given prompt.

Underpinning these improvements, GPT-4.5 is powered by an AI development technique known as unsupervised learning, which plays a key role in its enhanced capabilities.

Many AI models are trained using a method known as supervised learning, which involves providing the developing large language model with a dataset where each piece of information is paired with an explanation. This elaboration, or metadata, helps guide the model’s learning process. In contrast, unsupervised learning eliminates the need for metadata, requiring the model to independently analyze and interpret the training dataset.

OpenAI states that it leveraged more infrastructure and training data for GPT-4.5 than in its previous unsupervised learning projects. As part of this effort, the company developed “scalable techniques that enable training larger and more powerful models with data derived from smaller models.” This suggests that OpenAI may have used earlier language models, such as its reasoning-focused o3 algorithm, to generate training data for GPT-4.5.

While GPT-4.5 is not specifically designed for reasoning tasks, it is still capable of handling complex challenges. In an internal evaluation by OpenAI, the model achieved a higher score on the SWE-Lancer Diamond programming challenge set than the company’s reasoning-optimized o3-mini-high model. Additionally, GPT-4.5 successfully answered just over one-third of the questions in a 2024 qualifying exam for the U.S. Math Olympiad.

The model is accessible to developers via an application programming interface (API). OpenAI has priced one million input tokens at 30 times the cost of GPT-4o and one million output tokens at 15 times the cost. Due to these pricing considerations, the company is currently “evaluating whether to continue serving it in the API long-term as we balance supporting current capabilities with building future models.”

GPT-4.5 is also being gradually introduced to ChatGPT. Initially, it is available exclusively to users subscribed to the top-tier Pro plan. OpenAI plans to extend access to the Plus and Team plans next week, followed by a rollout to Enterprise and Edu tier users a week later.