OpenAI Releases Next-Generation Reasoning Model o1: Enhanced Math, Science, and Coding Capabilities
OpenAI CEO Sam Altman announced the launch of the o1 series of models, the company’s next-generation AI reasoning system, which includes o1-preview and o1-mini. Through a built-in chain-of-thought mechanism, the model excels in fields like math, science, and coding, and has been integrated into the ChatGPT platform for users to experience.
Core Reasoning Abilities of the o1 Model
Unlike traditional language models, the o1 model automatically generates an internal chain of reasoning when solving problems, simulating a human’s step-by-step thought process. According to OpenAI’s official blog, o1-preview scored 83% on the International Mathematical Olympiad qualifiers, far surpassing the previous GPT-4o’s 13%. On the Codeforces coding benchmark, o1 reached the 89th percentile, and achieved 74% accuracy on the GPQA science task. These results were cross-verified by reports from TechCrunch and Ars Technica, consistent with the original information from The Verge.
o1-mini, as a lightweight version, focuses on coding and math with lower training costs and faster response times, making it suitable for high-frequency use cases. OpenAI emphasizes that this model series undergoes more stringent safety testing, including external red teaming, to mitigate potential risks.
ChatGPT Integration and Access
The o1 model is now available in ChatGPT, accessible to Plus, Team, Enterprise, and Edu subscribers with daily limits. It is not yet available to free users, but OpenAI plans a gradual rollout. API access will be opened soon, with o1-preview priced at $15 per million input tokens and $60 per million output tokens. The more economical o1-mini will cost $3 per million input tokens and $12 per million output tokens. This information was confirmed by the official OpenAI website and Reuters.
Future Open-Source Plans include the gradual release of system cards, a reasoning trace visualizer, and model distillation components to support further development by the research community.
Performance Benchmarks and Application Scenarios
In multiple leading benchmarks, o1-preview surpasses Claude 3.5 Sonnet and Gemini 1.5 Pro. For example, it scored 74.3% on the AIME 2024 math competition and showed a 2x improvement on the ARC-AGI task. o1-mini excels in programming tasks, achieving an ELO rating of 1891, placing it among the top 500 developers. This data is sourced from OpenAI’s system card and analysis by VentureBeat.
OpenAI states that o1 is suitable for scientific research, education, and software development tools, helping users tackle multi-step complex problems. The company will continue to iterate and expects o1 to play a role in even more domains.
