OpenAI has recently introduced the o1-preview, a new series of advanced reasoning models designed to tackle more complex problems across various fields like science, math, and coding. Unlike previous models, these new AI systems are built to "think" more deliberately, spending extra time reasoning through difficult tasks before generating answers. The idea is to mimic human problem-solving abilities, allowing these models to refine their strategies and learn from their mistakes.

These enhanced capabilities are aimed at solving more challenging tasks that previous AI models might struggle with. The o1 series is particularly adept at problem-solving in high-level disciplines such as physics, chemistry, biology, and mathematics. In early tests, for instance, o1 models achieved impressive results. When tested on International Mathematics Olympiad (IMO) questions, GPT-4o, a predecessor to the o1 models, only solved 13% of problems, while the reasoning-focused o1-preview model solved 83% of them. This marks a significant leap in performance for solving difficult problems in coding and mathematics.

OpenAI's new models, o1-preview and o1-mini, are already available for use in ChatGPT and through the API, though they are in the early stages of development. Currently, the preview model doesn’t include some of the more common ChatGPT features like web browsing, file uploads, or image uploads, but OpenAI plans to add these features as the models mature. The o1 series excels in deep reasoning tasks, with the models trained to evaluate multiple strategies, recognize errors, and apply corrections in real time—skills that make them particularly useful for high-level scientific research and coding.

A notable update to the o1 series is its significant improvements in AI safety. OpenAI has introduced new safety protocols alongside the launch of these models. These safety measures ensure that the o1 models adhere strictly to ethical and safety guidelines. In comparison to earlier models like GPT-4o, the o1-preview scored 84 out of 100 on a "jailbreaking" test, a measure of how well the model can resist manipulation attempts. This improvement shows the new models’ enhanced ability to comply with safety protocols and their robustness in real-world applications.

The o1-preview and the more affordable o1-mini are aimed at different user needs. While o1-preview is suited for those needing high-level reasoning across a wide range of subjects, o1-mini is a smaller, faster model focused on coding tasks. The o1-mini is priced 80% cheaper than the o1-preview, making it an excellent option for developers and businesses that require reasoning capabilities but not necessarily broad world knowledge. For instance, developers can use o1-mini to build and debug complex code with efficiency without the cost of using the full o1-preview model.

As part of the ongoing rollout, OpenAI has made o1 models available to ChatGPT Plus and Team users, and API access has already begun for developers in the qualified usage tiers. ChatGPT Enterprise and Education users can expect to gain access soon. Users will be able to select between o1-preview and o1-mini manually within the ChatGPT interface, though OpenAI plans to optimize this by automatically selecting the right model based on the input prompt in the future.

Looking forward, OpenAI is committed to continually improving and expanding the o1 series. The company has outlined plans to add further updates, including features like browsing, file and image uploads, and other enhancements to make the models more useful. As OpenAI continues to refine the o1 models, it promises to bring even greater advancements to the field of AI reasoning.