SAN FRANCISCO — In a major step forward for AI development, OpenAI has introduced a new series of models designed to tackle more challenging tasks, such as those found in science, mathematics, and coding. These models, which emphasize more thoughtful processing before responding, are now available in ChatGPT and through the API. The models, known as OpenAI o1, are expected to bring significant advancements in AI capabilities and provide a foundation for the next generation of intelligent solutions.

At its core, the OpenAI o1 series is engineered to enhance the model’s reasoning abilities, allowing it to more effectively navigate and solve complex problems. During testing, the models outperformed previous versions in benchmark tasks, excelling particularly in math and coding. The models demonstrated a substantial improvement in solving International Mathematics Olympiad (IMO) problems, achieving an 83% success rate compared to only 13% for earlier versions. Additionally, in coding competitions, the models ranked within the 89th percentile.

While the o1 models are still in early development and may not yet include features like web browsing or file uploading, they represent a substantial leap forward in AI reasoning. OpenAI’s new models are expected to be invaluable for professionals working in areas that require advanced problem-solving, such as healthcare research, quantum optics, and software development.

One notable feature of these new models is their ability to reason through complex tasks, making them more aligned with the way humans approach problem-solving. Through extensive training, the models refine their approach, experimenting with different strategies and learning from their mistakes. This makes them particularly useful for industries that require intricate calculations or precise problem-solving, such as physics, chemistry, and biology.

In terms of safety, OpenAI has integrated a new safety training approach into the o1 models, which allows them to better adhere to safety and alignment guidelines. The models’ ability to reason through safety rules improves their effectiveness in preventing security breaches. OpenAI is also collaborating with the U.S. and U.K. AI Safety Institutes to ensure that the models meet stringent safety standards before their public release.

For developers and businesses, OpenAI is offering two versions of the o1 models: o1-preview and the more cost-effective o1-mini. The o1-mini model, while slightly smaller in scope, excels at generating and debugging complex code and is significantly cheaper than its counterpart. Both models are now accessible to ChatGPT Plus and Team users, with future availability expanding to educational and enterprise users. Developers who qualify for API usage tier 5 can begin prototyping with the models today.

OpenAI also plans to expand the functionality of the o1 series, adding features such as browsing capabilities, file uploading, and more. These updates, alongside further advancements in AI technology, will continue to shape the future of generative AI and its applications in solving real-world challenges.