OpenAI Launches o1: New AI Models for Complex Problem Solving

OpenAI has launched a new series of AI models designed to tackle complex reasoning tasks. These models, part of the new OpenAI o1 series, are engineered to "think" more deeply before responding, making them highly effective at solving difficult problems in science, coding, and math. The first models in the series, o1-preview and o1-mini, are available starting today in ChatGPT and via the API.

How OpenAI o1 Models Work

The o1 models are trained to spend more time processing information before providing answers, mimicking how humans reason through challenges. These models refine their thought process, experiment with different strategies, and learn from mistakes.

In recent tests, the next o1 model performed on par with PhD students on benchmark tasks in physics, chemistry, and biology. It also showed significant improvement in math and coding, solving 83% of problems in the International Mathematics Olympiad (IMO), compared to just 13% by GPT-4o. In coding contests, the o1 model ranked in the 89th percentile in Codeforces competitions.

Early Access and Capabilities

While o1 represents a major advancement in AI reasoning, it currently lacks some features available in ChatGPT, such as browsing the web and handling multimedia files. For routine tasks, GPT-4o may still be the preferred model, but o1 excels in complex reasoning.

To address safety concerns, OpenAI has developed a new training approach that enhances the model’s ability to follow safety guidelines. By reasoning through our safety rules in context, the model can apply them with greater accuracy and effectiveness. In a stringent "jailbreaking" test, o1-preview scored 84 out of 100 in resisting attempts to bypass safety rules, compared to GPT-4o’s score of 22.

OpenAI’s Commitment to AI Safety and Collaboration

As part of its broader focus on safety, OpenAI has ramped up internal governance, testing, and collaboration with federal agencies. This involves rigorous testing and evaluation through our Preparedness Framework, top-tier red teaming, and board-level reviews, including oversight by our Safety & Security Committee. Notably, OpenAI has formalized agreements with the U.S. and U.K. AI Safety Institutes, granting them early access to a research version of o1 to evaluate its safety. This was a crucial first step in our partnership, helping to establish a process for researching, evaluating, and testing future models both before and after their public release.

The Potential of o1 in Complex Fields

The o1 models are particularly useful if you’re tackling complex problems in science, coding, math, healthcare, physics, and software development. For example, o1 can assist healthcare researchers with annotating cell sequencing data, help physicists to generate complex mathematical formulas required for quantum optics, or help developers execute multi-step workflows with precision.

Introducing OpenAI o1-mini

In addition to the full o1 model, OpenAI is launching o1-mini, a smaller, faster, and more cost-effective reasoning model. While it’s not as powerful as o1-preview, it excels at coding tasks and is 80% cheaper, making it ideal for applications that require reasoning without the need for extensive world knowledge.

How to Access OpenAI o1

Starting today, ChatGPT Plus and Team users can manually select o1-preview and o1-mini in the model picker drop down menu. At launch, the weekly message limits will be 30 for o1-preview and 50 for o1-mini. OpenAI is working to raise these limits and allow ChatGPT to automatically select the appropriate model for each prompt. Both models will be available to ChatGPT Enterprise and Edu users next week. API access is currently available to developers in tier 5, with rate limits set at 20 RPM. The API for these models currently lacks features such as function calling, streaming, and support for system messages. To begin, you can refer to the API documentation.

OpenAI also plans to make o1-mini available to all ChatGPT Free users.

Looking Forward

OpenAI plans to bring more features to the o1 series, such as web browsing, file and image uploads, and broader functionality to support everyday use. Alongside this series, OpenAI will continue to develop its GPT models to enhance their capabilities in tandem with the new o1 family.