For certain developers, OpenAI's o1 reasoning model is available through its API

OpenAI Introduces o1 “Reasoning” AI Model to Select Developers via API

OpenAI is beginning the rollout of its o1 "reasoning" AI model through its API, starting with a limited group of developers. From Tuesday, developers in OpenAI’s “tier 5” usage category will gain access. To qualify for this category, developers must have an account that’s at least 30 days old since their first payment and have spent a minimum of $1,000 with OpenAI.

The o1 model replaces the earlier o1-preview model available via the API.

Unlike most AI systems, reasoning models like o1 are designed to self-validate their responses, helping them avoid common errors seen in traditional models. However, this comes at a cost — they’re slower to generate responses and are notably expensive due to high computational demands. OpenAI charges $15 to process approximately 750,000 words and $60 to generate the same amount, making it six times more costly than GPT-4o, the company’s latest standard model.

The o1 API offers significant improvements over o1-preview, with enhanced customizability. New features include function calling (which links the model to external data), developer messages (for instructing tone and style), and image analysis. Developers can also adjust the “reasoning_effort” parameter to control how much time the model spends refining its answers.

OpenAI described this API version of o1, labeled “o1-2024-12-17,” as a post-trained iteration with improvements informed by user feedback. Compared to the o1 model recently launched in ChatGPT, the updated version offers better model behavior and more precise responses, particularly for programming and business queries. OpenAI is expanding access in phases while gradually increasing rate limits for users.

In related announcements, OpenAI unveiled updates to its GPT-4o and GPT-4o mini models, which are part of its Realtime API. These updates, named “gpt-4o-realtime-preview-2024-12-17” and “gpt-4o-mini-realtime-preview-2024-12-17,” improve data efficiency, reliability, and cost-effectiveness.

The Realtime API, still in beta, gained new features such as concurrent out-of-band responses for background tasks like content moderation. It also supports WebRTC, an open standard for real-time voice applications used in browsers, smartphones, and IoT devices. OpenAI recently hired WebRTC creator Justin Uberti to strengthen this capability, ensuring seamless interactions even under fluctuating network conditions.

Finally, OpenAI introduced preference fine-tuning to its fine-tuning API, allowing developers to train models to prioritize preferred answers over less desirable ones. The company also launched an early access beta for official SDKs in Go and Java.