What is GPT-4.5
GPT-4.5 is the latest large chat model launched by OpenAI, and it is currently the largest and most powerful chat model. Based on scaled unsupervised learning, it improves pattern recognition, breadth of knowledge, and creative generation capabilities, reduces hallucinations, and makes conversations more natural and more consistent with human communication habits. The core advantages of GPT-4.5 lie in its broad knowledge base, higher "emotional intelligence," and precise understanding of user intent, making it suitable for scenarios such as writing, programming, solving practical problems, and creative design. GPT-4.5 supports web search, file upload, and image upload features. GPT-4.5 has undergone rigorous safety testing, improving the model's safety and reliability.
Main features of GPT-4.5
- Natural language conversation: Provides a more natural and warmer conversation experience, understands user intent, and gives answers that fit their needs.
- Writing assistance: Helps users with creative writing, copywriting, article polishing, and more, providing more coherent and more creative text.
- Programming support: Supports code generation, debugging, and optimization, and handles complex programming tasks and multi-step coding workflows.
- Knowledge Q&A: Has a broader knowledge base, answers questions in various fields, and reduces hallucinations.
- Multilingual support: Supports conversation and writing in multiple languages, with significant improvements in some rare languages (such as Swahili and Yoruba).
- Emotion understanding: Better captures users' emotions and needs, and provides more thoughtful suggestions or responses.
- Supports multiple input methods: Supports text, file, and image input. GPT-4.5 supports web search to obtain the latest information and provide more timely answers.
- API integration: Supports the Chat Completions API, Assistants API, and Batch API to integrate GPT-4.5 into various applications.
Technical principles of GPT-4.5
- Scaling of unsupervised learning: Improves its performance based on large-scale unsupervised learning. Unsupervised learning means the model is trained on a large amount of unlabeled data to learn language patterns and structures. GPT-4.5 further improves the model's understanding of the world based on expanded computing resources, data volume, and innovations in architecture and optimization.
- Deep World Knowledge: Trained on the Microsoft Azure AI supercomputer, it integrates broader knowledge and a deeper understanding of the world. This makes the model more reliable when handling various topics and reduces hallucinations.
- Human collaborative training: GPT-4.5 is based on new scalable technology, using data generated by small models to train larger models. It improves the model's controllability, understanding of nuances, and natural conversation ability. Based on human preference tests, GPT-4.5 outperforms previous-generation models in conversational naturalness and understanding ability.
- Enhancement of emotional intelligence (EQ): GPT-4.5 improves its understanding of human intent and emotional intelligence, better capturing implicit cues and expectations in conversations. It shows stronger empathy and creativity in conversations and generates content that is more consistent with human communication habits.
- Safety and alignment: During training, it combines traditional supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) methods to improve the model's safety and ensure that behavior aligns with human values.
Performance of GPT-4.5
- Knowledge Q&A (SimpleQA): GPT-4.5 achieved an accuracy of 62.5% on the SimpleQA dataset, leading other OpenAI models.
- Hallucination rate: In the SimpleQA hallucination rate evaluation, GPT-4.5's hallucination rate was 37.1% (lower is better), significantly lower than other models.
- PersonQA dataset: On the PersonQA dataset, GPT-4.5 achieved an accuracy of 0.78, clearly outperforming GPT-4o (0.28) and o1 (0.55).
- Multilingual performance: GPT-4.5 surpasses in multilingual evaluation (MMLU)GPT-4o, with notable improvements especially in rare languages (such as Swahili and Yoruba).
- Safety and bias detection: GPT-4.5 performs well in refusing harmful content, with a slightly higher over-refusal rate than previous-generation models.
- Programming and task execution: In programming tasks, GPT-4.5 has improved performance in code generation and repair tasks, especially in Agentic Tasks evaluations where it can complete some complex tasks, such as executing terminal operations in a Linux + Python environment, automatically downloading and running programs, etc.
- Human preference tests: In comparative evaluations with human testers, GPT-4.5 has a higher win rate (human preference tests) compared to GPT-4o, including but not limited to creative intelligence (56.8%), professional questions (63.2%), and everyday questions (57.0%).
How to use GPT-4.5
- ChatGPT Pro users: Starting February 28, ChatGPT Pro users can use the research preview of GPT-4.5.
- Plus and Team users: In the first week of March, GPT-4.5 will gradually be made available to ChatGPT Plus and Team users.
- Enterprise and Edu users: In the second week of March, GPT-4.5 will be made available to Enterprise and Edu users.
GPT-4.5 project official website
- Official website address：https://openai.com/index/introducing-gpt-4-5
GPT-4.5 product pricing
- Input cost: $75/million tokens
- Output cost: $150/million tokens
GPT-4.5 application scenarios
- Writing assistance: Helps generate text and polish content, suitable for writing, copywriting, etc.
- Programming support: Provides code suggestions and debugging help, improving programming efficiency.
- Learning tutoring: Answers academic questions and assists with review and exam preparation, suitable for students and educational scenarios.
- Communication and collaboration: Natural conversation, suitable for customer service, team collaboration, and cross-cultural communication.
- Task automation: Handles complex tasks and supports multi-step workflows, improving efficiency.