OpenAI Unveils GPT-4.5, Its Most Capable General-Purpose Model Yet

OpenAI has unveiled GPT-4.5, positioning it as the most expansive general-purpose large language model in its portfolio. The model joins two distinct collections of LLMs maintained by the ChatGPT creator: one optimized for reasoning-intensive work like code generation, and another focused on creative and knowledge-based text generation tasks. GPT-4.5 belongs to the latter group.

The new model demonstrates marked improvements over its predecessor, GPT-4o, particularly in how it processes user instructions. According to OpenAI's announcement, GPT-4.5 excels at recognizing "subtle cues or implicit expectations" in prompts, enabling it to produce outputs that more precisely match what users are asking for.

Another significant advancement involves emotional intelligence capabilities. GPT-4.5 has gained the ability to better gauge user intent in ways that previous models could not, such as determining whether someone wants a concise answer or a comprehensive one. This refinement allows the model to handle certain applications with greater accuracy.

Training Approach and Performance

GPT-4.5 relies on unsupervised learning, a training methodology that differs from the supervised approach used by many AI systems. While supervised learning pairs training data with explanatory metadata to guide the model, unsupervised learning requires the model to extract patterns from raw data without such guidance.

OpenAI invested substantially more computational resources and training data in this project compared to earlier unsupervised learning efforts. The company developed "scalable techniques that enable training larger and more powerful models with data derived from smaller models." This suggests that OpenAI leveraged earlier systems, including its reasoning-focused o3 algorithm, to generate training material for GPT-4.5.

Despite lacking specialized reasoning optimization, GPT-4.5 demonstrates strong performance across demanding tasks. Internal testing showed the model outperforming OpenAI's o3-mini-high reasoning model on the SWE-Lancer Diamond programming challenge set. The model also solved approximately one-third of questions from a 2024 U.S. Math Olympiad qualifying exam.

Availability and Pricing

Developers can access GPT-4.5 through an API, though pricing reflects its advanced capabilities. Input tokens cost 30 times more than GPT-4o, while output tokens carry a 15-fold premium. Given these costs, OpenAI stated it is "evaluating whether to continue serving it in the API long-term as we balance supporting current capabilities with building future models."

The model is rolling out across ChatGPT's subscription tiers in phases. Initially, GPT-4.5 is accessible exclusively through the Pro plan. The Plus and Team plans will gain access the following week, with Enterprise and Edu subscribers receiving availability the week after that.