SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

o1-pro - OpenAI's upgraded reasoning model

Author: AI小集 · Original source: AI工具集

What is o1-pro

o1-pro is an upgraded version of the o1 series officially released by OpenAI. o1-pro is currently OpenAI's most powerful reasoning model. Its core advantage lies in significantly enhanced computational capability, enabling it to better handle complex problems and provide more consistent, higher-quality responses. o1-pro is only available to specific developers (Tier 1–5). It supports vision, function calling, and structured outputs, and is compatible with the Responses and Batch APIs. In programming and mathematics, its performance is only slightly better than standard o1, with improved reliability.

o1-pro

Main features of o1-pro

  • Powerful reasoning capability: o1-pro uses more computational resources, can think more deeply, provides more accurate and reliable responses, and performs outstandingly when solving complex problems.
  • Supports multimodal input: Supports text and image input; output currently supports text only.
  • Structured outputs: Supports structured outputs, ensuring that output content conforms to a specific data format.
  • Function calling: Supports function calling and can connect to external data sources.
  • High context length and output limits: Has a 200,000 token context window and can generate up to 100,000 tokens per request.
  • Compatible with multiple APIs: Compatible with the Responses and Batch APIs, making it convenient for developers to use in different scenarios.

Technical principles of o1-pro

  • Reinforcement Learning (RL): o1-pro uses reinforcement learning to optimize its reasoning process. Through a process reward model (PRM), the model can receive immediate feedback while generating reasoning steps and gradually improve its reasoning strategy.
  • Process Reward Model (PRM): PRM provides reward signals for each reasoning step, rather than only based on the final result. This enables the model to better understand and optimize the reasoning process.
  • Monte Carlo Tree Search (MCTS): o1-pro uses Monte Carlo Tree Search (MCTS) during reasoning to explore different reasoning paths. MCTS helps the model choose the optimal path by simulating multiple possible reasoning steps. The method is similar to the tree search in AlphaGo, enabling the model to find better solutions to complex problems.
  • Self-Consistency mechanism: o1-pro adopts a self-consistency mechanism during the reasoning stage, improving the accuracy and reliability of reasoning by generating multiple reasoning paths and performing majority voting. It can effectively reduce error accumulation caused by a single reasoning path.
  • Synthetic data generation: To train o1-pro, OpenAI developed a system called "Berry Training," which generates large amounts of synthetic data through Monte Carlo trees. The data is filtered through functional verifiers and an optimization reward model to ensure the quality of the training data.
  • Test-Time Compute: o1-pro can use more computational resources during inference, improving the accuracy and depth of reasoning by increasing the amount of computation at test time. This enables the model to think more deeply when handling complex problems.

Project address of o1-pro

Pricing of o1-pro

  • Input price: $150 per 1 million tokens (about 750,000 English words).
  • Output price: $600 per 1 million tokens.
  • Batch API pricing: Input price is $75 per million tokens, output price is $300 per million tokens.
  • Other information: The price of o1-pro is 10 times that of the regular o1 model, and 2 times the input price of GPT-4.5. o1-pro is currently only available to specific developers (Tier 1–5), and developers need to spend at least $5 on API services to use it.

o1-pro performance testing

  • Mathematical reasoning: When handling doctoral-level scientific problems, o1-pro's accuracy improved to 79.3%; when solving American Invitational Mathematics Examination (AIME) problems, its accuracy reached 85.8%.
  • Programming ability: In the International Olympiad in Informatics (IOI), o1-pro's performance is significantly better than the regular o1 model.
  • Multimodal input: o1-pro supports image and text input and can handle complex multimodal problems. For example, when handling the heat dissipation design problem for a space data center, o1-pro can provide a detailed solution based on a hand-drawn sketch and the problem description.
  • Compared with the regular o1 model: o1-pro's performance on math and programming tasks improved by 7.5% and 2 times, respectively.
  • Compared with GPT-4.5: o1-pro's input and output prices are 2 times and 10 times those of GPT-4.5, respectively. In terms of performance, o1-pro is significantly better than GPT-4.5 in multiple benchmark tests.

o1-pro application scenarios

  • Interdisciplinary research: Supports multimodal input and can handle complex tasks combining images and text, such as analyzing the heat dissipation design problem for a space data center.
  • Code generation and optimization: o1-pro can generate high-quality code based on flowcharts, supports multiple programming languages and frameworks, and is suitable for complex coding tasks.
  • System architecture design: Provides in-depth code analysis and system architecture suggestions to help developers optimize software design.
  • Visual reasoning: Can analyze and reason about images, for example, performing complex calculations based on hand-drawn sketches.
  • Academic writing assistance: Can generate high-quality academic papers, debate scripts, poetry, etc., suitable for academic writing and creative writing.