GPT-4o's new version suddenly went online, with capabilities improved across the board, directly jumping to first place in the ZeroEval benchmark test.
Input and output are reduced by 50% and 33% respectively, which is effectively equivalent to an API price reduction. Token output has been expanded to 16k, previously supporting 4k.
OpenAI has introduced structured output in the API, using JSON schemas to ensure model output conforms to developer-defined structures, making the model more reliable and secure. The latest model version "GPT-4o-2024-08-06" scored 100% in JSON schema evaluation, a significant improvement from 40% in the June version last year.
In terms of code editing, the latest version scores the same as the previous version on benchmarks, but at half the cost. Additionally, the new version is OpenAI's cheapest model for image input.
The API supports structured output mainly in two forms:
- Function calls: Structured output is invoked by setting strict:true in the function definition.
- New option for response_format parameter: Developers can provide JSON schemas through json_schema.
OpenAI has updated Python and Node SDKs to support local structured output functionality. Structured output can also be used for dynamically generating user interfaces, separating final answers from reasoning processes, extracting structured data from unstructured data, and various other use cases.
Regarding pricing, the input price is $2.5 per million tokens, and the output price is $10 per million tokens. Since the new model is more efficient in input and output, it effectively amounts to a price reduction. Some users have found that the new model's output is not as verbose as the previous version, making costs more controllable.