OpenAI on Thursday introduced GPT-4.5, its latest and most advanced chat model, now available as a research preview. The company described GPT-4.5 as a significant step forward in scaling unsupervised learning, enabling the model to recognize patterns, draw connections, and generate creative insights more effectively than its predecessors.
GPT-4.5: Key Features and Improvements
GPT-4.5 builds on advancements in pre-training and post-training, offering a broader knowledge base, improved ability to follow user intent, and enhanced emotional intelligence (EQ).
Early tests suggest the model feels “more natural” in conversations and is expected to “hallucinate less,” making it more reliable for tasks like writing, programming, and problem-solving.
The model was trained on Microsoft Azure AI supercomputers, leveraging increased compute power, data, and architectural innovations. This has resulted in deeper world knowledge, reduced hallucinations, and greater reliability across diverse topics.
Scaling Unsupervised Learning and Reasoning
OpenAI highlighted two key approaches to advancing AI capabilities: unsupervised learning and reasoning.
- Scaling Reasoning: Models like OpenAI o1 and o3-mini focus on teaching AI to think through problems step by step, excelling in STEM and logic-based tasks.
- Unsupervised Learning: This approach enhances the model’s intuition and world understanding, as seen in GPT-4.5.
By combining these paradigms, GPT-4.5 achieves a balance of creativity and factual accuracy.
Performance and Comparisons
OpenAI compared GPT-4.5 with earlier models like GPT-4T and GPT-4o using the SimpleQA test, which evaluates factual accuracy and hallucination rates:
- Accuracy: GPT-4.5 scored 62.5%, outperforming GPT-4o (38.2%), OpenAI o1 (47%), and o3-mini (15%).
- Hallucination Rate: GPT-4.5 had a lower rate at 37.1%, compared to GPT-4o (61.8%) and o3-mini (80.3%).
Human testers also preferred GPT-4.5’s responses across various categories:
- Everyday Queries: 57% preferred GPT-4.5.
- Professional Queries: 63.2% favored GPT-4.5.
- Creative Tasks: 56.8% found GPT-4.5 more engaging.
Use Cases and Examples
GPT-4.5 excels in tasks requiring creativity, nuanced understanding, and collaboration. For instance, when asked about a painting depicting women setting their boats on fire, GPT-4.5 accurately identified Claude Lorrain’s The Trojan Women Setting Fire to Their Fleet, providing historical context and artistic details.
Safety and Training
OpenAI emphasized safety improvements in GPT-4.5, combining traditional supervised fine-tuning (SFT) with reinforcement learning from human feedback (RLHF). The model underwent rigorous safety testing under OpenAI’s Preparedness Framework, ensuring alignment with human values.
Availability
Starting February 27, 2025, GPT-4.5 will be available to ChatGPT Pro users on web, mobile, and desktop. It will roll out to Plus, Team, Enterprise, and Edu users in subsequent weeks.
The model supports search, file and image uploads, and canvas features but does not yet include multimodal capabilities like Voice Mode, video, and screen sharing in ChatGPT. The company stated that in the future, they will work to simplify the user experience so AI “just works” for users.
For developers, GPT-4.5 is available in the Chat Completions API, Assistants API, and Batch API. It supports features like function calling, structured outputs, streaming, and image inputs. However, due to its compute-intensive nature, GPT-4.5 is more expensive than GPT-4o, and OpenAI is evaluating its long-term availability in the API.
Future Directions
OpenAI believes reasoning will be a core capability of future models, complementing pre-training advancements. As models like GPT-4.5 grow smarter, they will serve as stronger foundations for reasoning and tool-using agents.
The company invites users to explore GPT-4.5 and provide feedback to help shape its future development.