OpenAI has released a research preview of GPT-4.5, its largest and most advanced model for chat to date. GPT-4.5 represents a significant step forward in both pre-training and post-training, with improvements in recognizing patterns, drawing connections, and generating creative insights without explicit reasoning. By scaling unsupervised learning, the model enhances its ability to process and understand complex queries more naturally.

Early testing indicates that interactions with GPT-4.5 feel more intuitive. The model boasts a broader knowledge base, improved ability to align with user intent, and greater emotional intelligence (EQ), making it highly effective for tasks such as writing enhancement, programming assistance, and problem-solving. Additionally, it is expected to exhibit fewer hallucinations than its predecessors.

GPT-4.5 has been released as a research preview to explore its strengths and limitations further. OpenAI remains committed to understanding its full capabilities and is eager to observe how users engage with the model in unexpected ways.

Scaling Unsupervised Learning

OpenAI continues to advance AI capabilities by scaling two key paradigms: unsupervised learning and reasoning. Unsupervised learning enhances the accuracy of the model’s understanding of the world, while reasoning enables AI to think methodically and construct logical chains of thought before responding.

GPT-4.5 is a prime example of scaling unsupervised learning by leveraging increased computational power and data, along with architectural and optimization innovations. The model was trained on Microsoft Azure AI supercomputers, resulting in a more knowledgeable and reliable system with reduced hallucinations across a wide range of topics.

Training for Human Collaboration

As AI models grow more powerful, it is increasingly important to enhance their ability to understand human needs and intent. GPT-4.5 incorporates new scalable training techniques that leverage data from smaller models to train larger ones more efficiently. These advancements improve its steerability, ability to interpret nuances, and capacity for natural conversation.

By combining deep world knowledge with improved collaborative abilities, GPT-4.5 integrates ideas more fluidly, making interactions feel warm and intuitive. It demonstrates a stronger grasp of human intent, better interpreting subtle cues and implicit expectations. Additionally, the model exhibits improved aesthetic intuition and creativity, excelling in writing and design-related tasks.

Advancements in Reasoning

Unlike specialized reasoning models such as OpenAI o1 and OpenAI o3-mini, GPT-4.5 does not think through responses before generating them. Instead, it serves as a general-purpose, inherently intelligent model. OpenAI envisions reasoning as a core capability of future models, with both pre-training and reasoning approaches complementing one another. As models like GPT-4.5 become more knowledgeable through pre-training, they will provide a stronger foundation for reasoning-driven AI and tool-using agents.

Safety Measures

Each advancement in AI capabilities presents an opportunity to improve model safety. GPT-4.5 was developed with enhanced supervision techniques alongside traditional supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), similar to those employed for GPT-4o.

To ensure safety, OpenAI conducted rigorous testing in accordance with its Preparedness Framework. Evaluations revealed that scaling the GPT paradigm contributed to improvements in model capabilities. Detailed results from these evaluations have been published in the accompanying system card.

Availability in ChatGPT

Starting immediately, ChatGPT Pro users can access GPT-4.5 through the model picker on web, mobile, and desktop platforms. The rollout will extend to Plus and Team users next week, followed by Enterprise and Edu users in the subsequent week.

GPT-4.5 supports real-time search for up-to-date information, file and image uploads, and canvas-based writing and coding tools. However, it currently does not support multimodal features such as Voice Mode, video, or screen sharing. OpenAI is working toward simplifying the user experience so AI functions seamlessly.

Availability in the API

OpenAI is also previewing GPT-4.5 in the Chat Completions API, Assistants API, and Batch API for developers on all paid usage tiers. The model supports key features such as function calling, structured outputs, streaming, and system messages. Additionally, it offers vision capabilities through image inputs.

Early testing suggests that GPT-4.5 is particularly useful for applications requiring high emotional intelligence and creativity, including writing assistance, communication, learning, coaching, and brainstorming. It also demonstrates strong capabilities in planning and execution, making it valuable for multi-step coding workflows and complex task automation.

Given that GPT-4.5 is a large and compute-intensive model, it is more costly than GPT-4o and is not intended as a direct replacement. OpenAI is evaluating whether to maintain long-term API support for GPT-4.5, balancing its availability with the development of future models. Feedback from users will play a crucial role in guiding this decision.

Conclusion

With each increase in computational power, AI models unlock novel capabilities. GPT-4.5 stands at the forefront of unsupervised learning, showcasing impressive improvements in knowledge depth and usability. OpenAI continues to be inspired by the creativity of the community in uncovering new applications. With GPT-4.5, the company invites users to explore the latest advancements in AI and contribute to the discovery of innovative use cases.