GPT-4o's New Version Stronger, Cheaper, and Structured Output
GPT-4o's new version suddenly went online, stronger and cheaper. The ability has been comprehensively enhanced, and it directly jumped to the first place in the ZeroEval benchmark test.
The input and output saved 50% and 33% respectively. It's almost like a price reduction for the API.
The token output has been extended to 16k, previously it supported 4k.
Even the question of which is greater, 9.11 or 9.9, can be smoothly solved.
This is because OpenAI introduced structured output in the API. By ensuring that the model output conforms to the structure defined by the developer through the JSON schema, it can make the model more reliable and secure.
The latest version of the model, GPT-4o-2024-08-06, scored 100% in the JSON schema assessment. In contrast, the version from last June scored less than 40%.
In terms of code editing, the benchmark score of the latest version is the same as that of the previous version, but now the cost is only half of the original.
Some people also found that this new version is the cheapest model of OpenAI in image input.
API supports structured output
At the OpenAIDevDay last year, the official officially introduced the JSON schema to GPT-4, which can help developers build applications.
Although the JSON schema can improve the reliability of the model, it cannot guarantee that the model generates results that conform to a specific schema.
Therefore, in this update, OpenAI introduces structured output in the API, which can ensure that the model-generated content completely matches the JSON schema provided by the developer.
The main forms are two:
1. Function call: By setting strict:true in the function definition to call structured output.
This feature is applicable to all model versions, including gpt-4-0613 and gpt-3.5-turbo-0613, etc. When structured output is enabled, the model output will match the provided tool definition.
2. New option of the response_format parameter.
Developers can provide a JSON schema through json_schema(response_format parameter's new option).
When the model is not calling a tool but responding to the user in a structured way, this is useful. This feature is applicable to gpt-4o-2024-08-06 and gpt-4o-mini-2024-07-18. When response_format is provided with strict:true, the model output will match the provided schema.
OpenAI updated the Python and NodeSDK to support the local structured output function.
In addition, security output is also emphasized. The new model will follow the existing security rules and introduce the refusal string value to help developers identify the situation where the model refuses unsafe requests.
Structured output can also be used for various use cases such as dynamically generating user interfaces, separating the final answer from the reasoning process, and extracting structured data from unstructured data.
Finally, in terms of price, The input price is $2.5 per 1 million tokens, and the output price is $10 per 1 million tokens.
Because the new model can save more on input and output, it's also a price reduction in a way.
Some people found that the output of this new model is not as verbose as the previous version, and the cost is more controllable.
