OpenAI brings an updated o1 model to its API, adding new features such as function calls, JSON-formatted structured output, and image analysis capabilities.
The new model, dubbed "o1-2024-12-17," significantly outperforms the previous "o1-preview" version. OpenAI reports the model achieves 96.4 percent accuracy on mathematical tasks, up from 85.5 percent, and 76.6 percent accuracy on programming tasks, a substantial increase from 52.3 percent.
Beyond raw performance improvements, the new o1 model uses about 60 percent fewer tokens for reasoning tasks compared to o1-preview, making it both faster and more cost-effective. Access to the new o1 model will be gradually rolled out, starting with level 5 API users.
The model is also available in ChatGPT, where it provides more comprehensive answers and fewer prompt rejections, with users consistently rating it higher than its predecessor, according to OpenAI.
Audio API gets major price cuts
OpenAI also announced significant price reductions for its audio features, cutting costs by 60 percent. Input tokens now cost $40 per million, while output tokens are priced at $80 per million. For cached audio inputs, the price drop is even more dramatic - an 87.5 percent reduction to $2.50 per million tokens.
The company also introduced a more affordable option called GPT-4o mini, available at one-tenth of the standard audio pricing: $10 per million input tokens and $20 per million output tokens. Developers can now create real-time voice applications more easily with new WebRTC integration, and the maximum session duration has been extended from 15 to 30 minutes.
New fine-tuning option for style and tonality
"Preference Fine-Tuning" is a new method for customizing AI models. Unlike traditional supervised fine-tuning, this approach learns from pairs of preferred and non-preferred responses, making it particularly effective for subjective tasks like creative writing and summaries where tone and style matter, OpenAI says.
The company also expanded its developer tools with new software development kits (SDKs) for Go and Java, complementing existing support for Python, Node.js, and .NET. Access to these features is being rolled out gradually, starting with level 5 users.
Developers can find detailed documentation for o1, the real-time API, WebRTC integration, and preference fine-tuning in the API documentation. OpenAI is also providing new software development kits (SDKs) for the Go and Java programming languages. These complement the existing SDKs for Python, Node.js and .NET.
