We’re excited to introduce GPT-4.5, our most advanced chat model yet, now available in research preview for Pro users and developers worldwide. This represents a significant step forward in scaling both pre-training and post-training techniques, pushing the boundaries of what’s possible with unsupervised learning.
Through advanced scaling of unsupervised learning, GPT-4.5 demonstrates remarkable improvements in pattern recognition, connection-making, and generating creative insights without explicit reasoning. Early testing reveals interactions feel more natural and intuitive, with users reporting a more human-like conversational experience.
The model’s enhanced capabilities stem from a broader knowledge base, improved understanding of user intent, and what we’re calling greater “EQ” – emotional intelligence that makes it particularly valuable for writing assistance, programming support, and solving practical problems. Notably, we’ve observed reduced hallucination rates compared to previous models.
Deeper World Knowledge and Improved Factuality
Our testing shows GPT-4.5 achieves impressive results on knowledge benchmarks:
- SimpleQA Accuracy: 62.5% (vs GPT-4o’s 38.2%)
- SimpleQA Hallucination Rate: 37.1% (vs GPT-4o’s 61.8%)
These improvements reflect the model’s enhanced understanding across a wide range of topics, making it more reliable for factual queries and complex knowledge tasks.
Enhanced Human Collaboration
GPT-4.5 demonstrates significant improvements in understanding human needs and intent. We’ve developed new scalable training techniques that enhance the model’s steerability, nuance comprehension, and natural conversation flow.
Human preference evaluations show GPT-4.5 winning against GPT-4o across various query types:
– Creative intelligence: 56.8% preference
– Professional queries: 63.2% preference
– Everyday queries: 57.0% preference
The model excels at interpreting subtle cues and implicit expectations, showing stronger aesthetic intuition and creativity that particularly benefits writing and design tasks.
Practical Applications and Emotional Intelligence
In real-world scenarios, GPT-4.5 demonstrates remarkable emotional intelligence. When a user shares they’re “going through a tough time after failing a test,” GPT-4.5 responds with empathy and offers conversation rather than immediately providing extensive advice. This contrasts with previous models that might default to listing solutions without first acknowledging the emotional context.
Technical Foundation and Safety
GPT-4.5 represents our continued advancement in scaling two complementary AI paradigms: unsupervised learning and reasoning. While reasoning models like OpenAI o1 excel at complex STEM problems through chain-of-thought processes, GPT-4.5 advances the unsupervised learning paradigm through compute scaling, architectural innovations, and optimization improvements.
Safety remains paramount. GPT-4.5 incorporates new supervision techniques combined with traditional SFT and RLHF methods, with extensive testing conducted in accordance with our Preparedness Framework.
Availability and Features
Starting today, ChatGPT Pro users can access GPT-4.5 through the model picker across web, mobile, and desktop platforms. The model supports:
– Latest information through search
– File and image uploads
– Canvas functionality for writing and code
– Vision capabilities through image inputs
While it doesn’t currently support Voice Mode, video, or screensharing, we’re committed to simplifying the AI experience to make it “just work” for users.
For developers, GPT-4.5 is available through our API with support for function calling, Structured Outputs, streaming, and system messages. Its higher emotional intelligence and creativity make it particularly valuable for writing assistance, communication tools, learning applications, coaching, and brainstorming.
We’re releasing GPT-4.5 as a research preview to better understand its capabilities and limitations, and we’re eager to see how the community discovers unexpected use cases. With every new order of magnitude in compute comes novel capabilities, and GPT-4.5 represents the frontier of what’s possible in unsupervised learning today.