Dec. 18, "OpenAI Entering the ninth day of its "12 Days" campaign, OpenAI today announced that its "reasoning" AI model, o1, has been officially approved for the API It is available to select developers and has been updated with a number of developer tools including GPT-4o, real-time APIs, and fine-tuning APIs.
It is reported that,The first developers with access to the o1 API are "Level 5" users of OpenAI.To reach this level, developers need to have spent at least $1,000 (currently around RMB 7,285) on the OpenAI platform and have had an account for more than 30 days since the first successful payment. To reach this level, developers need to have spent at least $1,000 cumulatively on the OpenAI platform (note: currently around Rs. 7,285) and have had an account for more than 30 days since the first successful payment. o1 API replaces the o1-preview model that was previously provided in the API.
Unlike most AI models, theo1, etc.inference modelCapable of effective self-fact-checkingOpenAI charges $15 per 750,000 words analyzed for o1 and $60 per 750,000 words generated, thus avoiding some common errors. The downside is that it is more computationally expensive and takes longer to reach a conclusion. openAI charges $15 per ~750,000 words analyzed and $60 per ~750,000 words generated for o1.Six times the size of its latest "non-inference" model, GPT-4o..
Compared to o1-preview, o1 in the API is more customizable, with new features such as function calls (allowing the model to connect to external data), developer messages (allowing the developer to direct the tone and style of the model), and image analysis. In addition to structured output, o1 provides an API parameter called "reasoning_effort" that allows the developer to control the amount of time the model spends "thinking" before answering a question.
OpenAI stated thatThe o1 in the API (and coming soon to its chatbot platform ChatGPT) is the "newly post-trained" versionThe version number is "o1-2024-12-17". Compared to the o1 model released in ChatGPT two weeks ago, the new version improves on several aspects of the model's behavior based on user feedback, but OpenAI did not disclose specific details.In a blog post, OpenAI said, "We are gradually rolling out access while working to expand access to cover more user levels and increase rate limits." In a note on its Web site, OpenAI said the newest o1 should provide "more comprehensive and accurate answers," especially on programming and business-related questions, and is less likely to reject requests incorrectly.
In other developer-related news.OpenAI also released new versions of the GPT-4o and GPT-4o mini modelsThe new models (gpt-4o-realtime-preview-2024-12-17, respectively) are available as part of its Realtime API. The Realtime API is OpenAI's API for building applications with low-latency, AI-generated voice responses. the new models ("gpt-4o-realtime-preview-2024-12-17" and "gpt-4o-mini-realtime-preview-2024-12-17", respectively) have greater data efficiency and reliability, and are more cost-effective to use. mini-realtime-preview-2024-12-17", respectively) have higher data efficiency and reliability, and are cheaper to use.
The Real-Time API is still in beta, but has gained several new features, such as concurrent out-of-band responses, which allow background tasks such as content review to run without interrupting interactions. The API also now supports WebRTC, the open standard for building real-time voice applications for browser-based clients, smartphones, and IoT devices. It's worth noting that OpenAI hired WebRTC creator Justin Uberti in early December. "Our WebRTC integration is designed to enable smooth and responsive interactions under real-world conditions, even when network quality is inconsistent," OpenAI wrote in a blog post. writes, "It can handle audio encoding, streaming, noise suppression, and congestion control."
As a final update, OpenAI has introduced preference fine-tuning to its fine-tuning API; preference fine-tuning compares pairs of model responses to "teach" the model to distinguish between preferred and "non-preferred" answers to a question. The company also launched an "early access" beta of its official software development kit for Go and Java.