OpenAI has just launched an exciting new series of reasoning models that promise to revolutionize problem-solving in fields like science, coding, and math. These innovative AI models, debuting on September 12, are designed to spend more time thinking before they provide responses, allowing them to tackle more complex tasks effectively.
The first models of this new series, available now through ChatGPT and the API, represent a significant leap forward in artificial intelligence. These models, trained to mimic the human thought process, refine their strategies and learning methods to produce smarter solutions. Early tests are impressive, with these models performing as well as PhD students on rigorous tasks in physics, chemistry, and biology. For instance, in a test for the International Mathematics Olympiad (IMO), the new reasoning model solved 83% of problems correctly, far surpassing the previous GPT-4o model, which managed only 13%.
In the realm of coding, these models excel as well, scoring in the 89th percentile in coding competitions on Codeforces. This makes them incredibly valuable tools for developers and researchers who need to generate and debug complex code efficiently.
This series, named OpenAI o1, includes a new safety training approach that ensures strict adherence to safety and alignment guidelines. This enhanced safety mechanism is particularly impressive; the latest model scored 84 out of 100 on one of the company’s toughest jailbreaking tests, compared to a score of 22 by the previous model.
OpenAI is committed to maintaining rigorous safety and governance standards. Collaborations with AI Safety Institutes in the U.S. and U.K. have been formalized, allowing these institutes to access research versions of the models. This partnership is crucial for establishing a robust process for evaluating and testing future models.
The o1 series’ capabilities are especially useful for tackling complex issues in various disciplines. Healthcare researchers can use it for annotating cell sequencing data, while physicists might find it invaluable for generating intricate mathematical formulas for quantum optics. Developers across all fields will benefit from its ability to build and execute multi-step workflows.
In addition to the standard o1 model, OpenAI is introducing the OpenAI o1-mini, a more affordable and faster variant specifically optimized for coding tasks. This model is 80% cheaper than the o1-preview without compromising on its reasoning abilities, making it a cost-effective choice for a wide range of applications.
Users of ChatGPT Plus and Team can start using these models immediately, with manual selection available in the model picker. Initial weekly message limits are set at 30 for o1-preview and 50 for o1-mini, with plans to increase these rates soon. Enterprise and educational users will also gain access next week.
Developers with access to API usage tier 5 can begin experimenting with both models starting today, with an initial rate limit of 20 requests per minute. Despite some initial limitations like lack of function calling and streaming support, these models present an excellent opportunity to prototype advanced solutions.
OpenAI also plans to roll out o1-mini access to all ChatGPT Free users soon. This early preview is just the beginning, with more updates and features like browsing and file uploading expected to follow. As OpenAI continues to develop both the GPT and o1 series, the future of AI looks incredibly promising.
