OpenAI recently launched a new version of its API, the o1 model, named "o1-2024-12-17". This version introduces several exciting new features, including intelligent function calls, support for structured output in JSON format, and image analysis capabilities.
According to OpenAI's report, the new model shows significant improvement in performance across multiple tasks, particularly in mathematics and programming. The newly released o1 model achieved an accuracy of 96.4% in mathematical tasks, a substantial increase from the 85.5% accuracy of the previous version "o1-preview".
At the same time, the accuracy in programming tasks has also increased from 52.3% to 76.6%. This series of improvements means that developers and users will achieve more accurate and efficient results when using this model, especially when dealing with complex problems, where accuracy is better assured.
In addition to the increase in accuracy, the new model also reduces token usage by approximately 60% when handling reasoning tasks. This not only enhances the model's operating speed but also lowers usage costs, making it more economical for developers to apply this model.
OpenAI plans to gradually roll out the new model to level 5 API users, ensuring that users can experience the convenience brought by this new technology.
Significant Price Reduction for Audio API
In terms of audio functionality, OpenAI has also made significant price adjustments. The cost of audio input tokens has been reduced by 60%, now costing only $40 per million input tokens, while output tokens are priced at $80. Furthermore, the cost for cached audio inputs has seen a more dramatic reduction, down to just $2.50 per million tokens.
Additionally, OpenAI has introduced a more affordable option called GPT-4o mini, priced at $10 per million input tokens and $20 for output tokens, providing more convenience for developing real-time voice applications.
New Styles and Fine-Tuning Options
OpenAI has also introduced a new method called "preference fine-tuning", which differs from traditional supervised fine-tuning. This method improves the model's performance on subjective tasks such as creative writing and summarization by learning users' preferences for different responses.
Moreover, OpenAI has expanded its developer tools, launching new software development kits (SDKs) for Go and Java, further enhancing the user experience for developers.
These new features will gradually be made available to developers, with OpenAI initially providing detailed documentation to level 5 users to help them better understand and utilize these new capabilities.
Key Points:
📊 The new o1 model's accuracy in mathematical and programming tasks has risen to 96.4% and 76.6%, respectively.
💰 OpenAI's audio functionality has seen significant price reductions, with audio input tokens dropping to $40 per million and output tokens to $80 per million.
🛠️ The newly introduced "preference fine-tuning" method and Go, Java SDKs empower developers to enhance customization capabilities.