Introducing OpenAI o1: A New Model for Solving Hard Problems

Update (September 17, 2024): The current rate limits for the o1-preview model are now set to 50 queries per week, while the o1-mini model is limited to 50 queries per day.

OpenAI has introduced a new series of AI models that are specifically designed to approach difficult problems with deeper reasoning. These models are capable of handling more complex tasks and outperform previous iterations in areas such as science, coding, and mathematics.

The first model in this series has been made available through ChatGPT and OpenAI's API. This early release marks the beginning of what will be an ongoing process of updates and enhancements. OpenAI is also working on the next model in this series, with evaluations already underway for its future improvements.

How It Works

The training of these models focuses on allowing them to take extra time in problem-solving, similar to how humans might approach challenging tasks. This deliberate thinking process enables the model to refine its strategies, correct its mistakes, and explore various solutions.

In rigorous testing, the upcoming update of this model demonstrated impressive performance, comparable to that of PhD students in difficult subjects like physics, chemistry, and biology. For instance, when tested on a qualifying exam for the International Mathematics Olympiad (IMO), the reasoning model scored 83%, a significant leap compared to GPT-4o, which solved only 13% of the problems. The model also excelled in coding, achieving results in the 89th percentile in Codeforces programming competitions. Further technical details can be found in OpenAI's research publications.

Currently, the o1-preview model lacks some features available in other models, such as the ability to browse the web or handle file and image uploads. While the more general-purpose GPT-4o might be preferable for routine tasks, the o1 model is a major advancement in AI reasoning capabilities, setting a new benchmark for complex problem-solving.

Safety Enhancements

OpenAI has developed a new approach to safety training for these models. By leveraging the models' advanced reasoning abilities, they can better adhere to safety guidelines. This reasoning-based safety framework helps the model apply rules effectively in real-world situations.

One important safety test involved evaluating how well the model adhered to guidelines when users attempted to bypass them—commonly known as "jailbreaking." In this test, GPT-4o scored 22 out of 100, while the o1-preview model scored an impressive 84. Further details are available in the system card and research posts shared by OpenAI.

In response to the enhanced capabilities of the o1 model, OpenAI has strengthened its internal governance, testing, and safety evaluations. Partnerships with AI Safety Institutes in the U.S. and U.K. have also been established, allowing for early access to research versions of the model to further develop safety protocols.

Who Should Use It?

The enhanced reasoning skills of the o1 series make it an excellent tool for professionals and researchers tackling complex challenges in fields like science, coding, and mathematics. For example, it can help healthcare researchers annotate cell sequencing data, aid physicists in generating complex formulas for quantum optics, or assist developers in building multi-step workflows for their projects.

OpenAI o1-mini

Alongside the full-scale o1-preview, OpenAI has also released a smaller, faster model called o1-mini. This version is specifically optimized for coding tasks and is 80% cheaper to run than o1-preview, making it a cost-effective option for developers. While it lacks some of the broader world knowledge of its larger counterpart, o1-mini excels at generating and debugging complex code.

How to Access OpenAI o1

The o1 models are available to ChatGPT Plus and Team users starting today. Both the o1-preview and o1-mini models can be manually selected in the model picker. At launch, the rate limits are set at 30 messages per week for o1-preview and 50 messages per week for o1-mini, with plans to increase these limits over time.

Starting next week, ChatGPT Enterprise and Edu users will also gain access to these models. Developers approved for API tier 5 usage can begin prototyping with both models right away, with a rate limit of 20 requests per minute (RPM). OpenAI plans to gradually increase API limits after further testing, though certain features, such as function calling and streaming, are not yet supported in the API.

OpenAI has also announced plans to extend access to the o1-mini model for all ChatGPT Free users in the near future.

Looking Ahead

This is an early preview of the o1 reasoning models, and OpenAI expects to roll out further updates and features, including browsing, file uploads, and more, to make these models even more versatile. Additionally, OpenAI will continue to develop both its GPT series and the new o1 series, pushing the boundaries of AI's problem-solving abilities.

Consult with our experts at Amity Solutions for additional information on Amity Chatbots here