OpenAI’s makes the full version of its o1 reasoning model available, but only to some developers – News

OpenAI said today it’s making the full version of its o1 reasoning model available to its most committed developer customers.

Today’s announcement came on the ninth day of its holiday-themed press blitz, known as “12 Days of OpenAI,” where the company said it’s rolling out access to the full o1 model for developers in the “Tier 5” category only. That means it’s restricted to developers who have had an account with OpenAI for at least one month, and who are spending at least $1,000 per month on its services.

Prior to today’s announcement, developers could only access the less powerful o1-preview model.

In addition to the restrictions on its use, the full version of the o1 reasoning model is very expensive, due to the enormous computing resources required to power it. According to the company, it will cost $15 for every 750,000 words analyzed, and $60 for every 750,000 words it generates. That makes it almost four times as expensive as the more widely used GPT-4o model.

Those who are prepared to pay the higher prices will at least get some new capabilities, as OpenAI has made a number of improvements compared to the preview iteration. For one thing, the full version of o1 is more customizable than the older version. There’s a new “reasoning_effort” parameter that dictates how long the model will ponder a specific question.

It also supports function calling, which means it can be connected to external data sources, plus developer messages and image analysis, which were not supported by the o1-preview model. Its latency has been reduced too, as it uses about 60% fewer reasoning tokens on average.

In other news, OpenAI said it’s incorporating the GPT-4o and 4o-mini models into its Realtime application programming interface, which is designed for low-latency vocal AI applications such as Advanced Voice Mode. The Realtime API also gets support for WebRTC, which is an open standard for developing vocal AI applications in web browsers. So that suggests we may well see a lot more websites trying to talk to their users in the coming months.

“Our WebRTC integration is designed to enable smooth and responsive interactions in real-world conditions, even with variable network quality,” OpenAI said in a blog post. “It handles audio encoding, streaming, noise suppression, and congestion control.”

Finally, there’s a new feature called “direct preference optimization” for developers who want to fine-tune their AI models. With its existing techniques for supervised fine-tuning, developers are required to provide examples of the input/output pairs they want to use to refine their models. But with this new feature, they can instead just provide two different responses and indicate which one is preferable to another.

According to the company, this will help optimize models to learn the difference between the user’s preferred and non-preferred answers, automatically detecting any changes in formatting, style guidelines or verbosity, and factor these into the new model.

The update is one of the most exciting so far in OpenAI’s 12-day media bonanza, following the launch of the Sora video generation model, a new Projects feature, and updates to Advanced Voice Mode, Canvas and Search.

Image: OpenAI