OpenAI has released its o1 model, claiming big improvements over the existing o1-preview (released in September this year), as well as introducing new SDKs for Java and Go, and improving its realtime API with WebRTC support

The o1 model is designed for complex multi-step tasks, at the expense of longer compute times, and requires developers to be on usage tier 5, reserved for organizations who have been spent over $1000 on API calls.

The company states that OpenAI o1 uses 60 percent fewer “reasoning tokens” than the preview, reducing latency and cost, and delivers more accurate results according to its benchmarks.

Features of o1 include the ability to define functions to augment the model, and the ability to accept visual inputs in order to reason over images. Using images has many applications in science and manufacturing. There is also a new API parameter called reasoning_effort, which controls how long the model thinks before responding, enabling users to control costs.

API access to OpenAI o1 Pro mode is not yet available but is promised soon. The Pro mode uses more compute for greater accuracy.

Realtime support in the OpenAI API has been improved with the addition of WebRTC support, as an alternative to coding directly for websockets. In this day 9 presentation, part of a 12-day marketing blitz, the company showed code to connect to an OpenAI model and conduct a voice chat using only a few lines of JavaScript code.

There are also two new models added to the realtime API, and prices are reduced. gpt-4o-realtime-preview-2024-12-17 has a 60 percent reduction in the audio token price, and gpt-4o-mini-realtime-preview-2024-12-17 is a cost-efficient small model priced at around 25 percent of the full version.

Another new feature is called preference fine-tuning or Direct Preference Optimization, and allows developers to customize the model’s responses according to preferences such as concise versus verbose responses. This supplements an existing approach called supervised fine-tuning.

OpenAI has also previewed new SDKs for Java and Go. Until now, the official OpenAI SDKs only offered Azure OpenAI libraries for Java and Go, stating that they are “compatible with both the OpenAI API and Azure OpenAI services.” That was not ideal, and the new SDKs are owned by OpenAI itself. 

The OpenAI news demonstrates the fast-moving nature of AI development. It is understandable if developers find it confusing. “What is this poorly documented reasoning_effort parameter?” asked one in response to the news.