OpenAI has launched a research preview of GPT-4.5, described as their most advanced and capable chat model to date. The model represents progress in scaling both pre-training and post-training methodologies, building upon unsupervised learning techniques.
Key Advancements
Enhanced Unsupervised Learning
GPT-4.5 advances capabilities through two complementary intelligence paradigms:
- Unsupervised learning - Enhances world model precision and intuitive understanding, the foundation of models like GPT-3.5, GPT-4, and now GPT-4.5
- Reasoning scaling - Enables models to develop thought chains before responding, crucial for complex STEM and logical challenges (exemplified by models like OpenAI o1 and o3-mini)
The model was developed using Microsoft Azure AI supercomputing infrastructure, incorporating architectural innovations and optimization improvements alongside expanded compute and data resources.
Performance Improvements
Factual Accuracy
- SimpleQA Accuracy: GPT-4.5 achieves 62.5%, surpassing GPT-4o (38.2%), OpenAI o1 (47%), and o3-mini (15%)
- Reduced Hallucinations: 37.1% hallucination rate compared to GPT-4o (61.8%), o1 (44%), and o3-mini (80.3%)
Human Preference Testing
GPT-4.5 demonstrated superior performance across multiple domains:
- Creative intelligence: 56.8% preference over GPT-4o
- Professional queries: 63.2% preference
- Everyday tasks: 57.0% preference
Enhanced Human Collaboration
OpenAI has implemented novel scalable techniques allowing larger model training using data from smaller models. These advancements improve:
- Model steerability and control
- Nuanced understanding of context
- Natural conversational flow
- Emotional intelligence and social awareness
- Creative and aesthetic capabilities
Benchmark Performance
GPT-4.5 shows significant improvements across academic benchmarks:
- GPQA (science): 71.4% (vs. GPT-4o: 53.6%)
- AIME '24 (math): 36.7% (vs. GPT-4o: 9.3%)
- MMMLU (multilingual): 85.1% (vs. GPT-4o: 81.5%)
- MMMU (multimodal): 74.4% (vs. GPT-4o: 69.1%)
- SWE-Bench Verified: 38.0% (vs. GPT-4o: 30.7%)
Availability and Access
ChatGPT Integration
- Immediate: Available to ChatGPT Pro subscribers
- Next week: Rollout to Plus and Team users
- Following week: Enterprise and Education users
The model includes support for:
- Web search with current information
- File and image processing
- Canvas functionality for writing and coding tasks
Current limitations include lack of Voice Mode, video, and screensharing capabilities.
API Access
Developers on all paid tiers can access GPT-4.5 through:
- Chat Completions API
- Assistants API
- Batch API
Supported features include:
- Function calling capabilities
- Structured Outputs
- Real-time streaming
- System message configuration
- Vision capabilities via image inputs
Safety Measures
OpenAI has implemented enhanced safety protocols including:
- Novel supervision techniques integrated with traditional supervised fine-tuning (SFT)
- Reinforcement learning from human feedback (RLHF) methodologies
- Comprehensive safety testing aligned with OpenAI's Preparedness Framework
- Detailed evaluation results published in an accompanying system card
Future Direction
While GPT-4.5 excels at pattern recognition and creative insights without explicit reasoning, OpenAI views the integration of reasoning capabilities as essential for future models. The organization believes that combining pre-training advances with reasoning capabilities will produce more capable foundation models for reasoning and tool-using agents.
Due to the model's substantial computational requirements and associated costs, OpenAI is evaluating its long-term API availability based on user feedback and demonstrated value for specific applications.