OpenAI has entered the ninth day of its ongoing “12 Days of OpenAI” event and has made several significant announcements regarding its developer tools. In a press release today, the company revealed that its reasoning AI model, termed o1, will now be accessible via API to select developers. This update is part of a broader enhancement that includes the release of GPT-4o, a real-time API, and a fine-tuning API.
The initial group granted access to the o1 API will consist of “Level 5” users on the OpenAI platform. To achieve this status, developers must have spent at least $1,000 on the platform, and their accounts must be over 30 days old since their first successful payment. The o1 API will replace the previous o1-preview model previously available to developers.
Differentiating itself from many AI models, o1 and similar reasoning models have the capability of self-fact-checking, which helps avoid common errors. However, this advanced functionality comes with higher computational costs and longer processing times. Specifically, OpenAI has priced the o1 services at $15 for analyzing approximately 750,000 words and $60 for generating the same amount, making it six times more expensive than its latest non-reasoning model, GPT-4o.
The new o1 API offers improved customization options compared to the previous version, including function calls that allow the model to connect to external data, developer messaging to guide the model’s tone and style, and enhanced image analysis capabilities. Additionally, the o1 API introduces a parameter called “reasoning_effort,” which allows developers to dictate how long the model should “think” before providing an answer.
OpenAI has labeled this iteration of the o1 API, versioned “o1-2024-12-17,” as a “recently retrained” model set to enhance user interactions based on feedback received since its launch in ChatGPT two weeks ago. The company stated, “We are gradually rolling out access while working to expand coverage to more user levels and improve rate limits,” emphasizing that the updated o1 aims to provide “more comprehensive and accurate answers,” particularly for programming and business inquiries.
In addition to the o1 updates, OpenAI has unveiled new versions of the GPT-4o and GPT-4o mini models, integral to its real-time API. This API is designed to support the creation of low-latency applications that require AI-generated voice responses. The updated versions, named “gpt-4o-realtime-preview-2024-12-17” and “gpt-4o-mini-realtime-preview-2024-12-17,” promise greater data efficiency, reliability, and reduced costs.
The real-time API is currently in testing but has introduced several new features, including concurrent out-of-band responses, which facilitate background tasks such as content moderation without interrupting user interaction. Additionally, it now supports WebRTC—an open standard for developing real-time voice applications for web browsers, smartphones, and IoT devices. OpenAI recently hired WebRTC creator Justin Uberti to enhance this integration, stating, “Our WebRTC integration aims to provide smooth and responsive interactions even under unstable network conditions.”
The final update from OpenAI includes the introduction of preference fine-tuning for its fine-tuning API, which helps the model identify preferred and non-preferred answers through paired comparisons. Furthermore, the company has launched an “early access” beta for software development kits (SDKs) tailored to Go and Java programming languages.
