AI that thinks: OpenAI introduces new 'o1' AI models that are better at reasoning, math, coding

OpenAI has launched a new series of AI models called OpenAI o1, which are designed to handle more difficult problems, especially in areas like science, coding, and maths. These models spend more time thinking before giving answers, similar to how a person would tackle tough questions. The first version of these models, o1-preview, is now available in ChatGPT and through an API, with more updates and improvements expected soon.

How it Works

The o1 models have been trained to think more carefully and deeply about problems. They learn from their mistakes and try different approaches when they get stuck. This training makes them better at reasoning through complex tasks. For example, the new model outperformed older ones in exams for physics, chemistry, and biology, and it solved 83 per cent of problems in a maths competition, compared to just 13 per cent by earlier models.

While o1-preview doesn’t yet have features like web browsing or uploading files, it represents a big improvement in handling complicated reasoning tasks.

OpenAI has developed new safety measures for these models. The o1-preview model is much better at following safety rules than previous models. For instance, it scored much higher on tests where users tried to trick it into breaking safety guidelines. OpenAI is also working with AI safety institutes in the US and UK to ensure the models are tested and evaluated thoroughly before wider release.

Who can benefit from OpenAI o1?

These models are particularly useful for people working on challenging problems in fields like healthcare research, physics, and coding. For example, they can help researchers analyse complex data or assist developers in creating more advanced software.

Introducing OpenAI o1-mini

Alongside the main o1 model, OpenAI is also releasing o1-mini, a smaller, faster, and cheaper version designed specifically for coding and STEM tasks. It performs nearly as well as the full o1 model but is 80 per cent cheaper, making it ideal for developers who need strong reasoning capabilities without the extra cost.

How to use OpenAI o1

If you use ChatGPT Plus or Team, you can access both the o1-preview and o1-mini models today, with some limits on how many messages you can send. Enterprise and Educational users will get access next week, while developers using the API can already start testing the models. OpenAI plans to bring o1-mini to all free ChatGPT users soon.

What’s Next?

This is just the beginning. OpenAI plans to keep improving these reasoning models and add more features like browsing and file uploads. They also plan to continue developing their existing GPT series alongside the new o1 models.