Mingmin from Aofei Temple

Quantum Bit | Public Account QbitAI

A new version of GPT-4o suddenly came online, which is stronger and cheaper.

With all-round improvements in capabilities, it directly jumped to first place in the ZeroEval benchmark test.

Input and output savings are 50% and 33% respectively.

Rounding it off means the API price has been reduced.

Token output has been expanded to 16k, previously supporting 4k.

The question of which is greater, 9.11 or 9.9, can also be solved smoothly.

This is becauseOpenAIIntroduced structured output to the API.

Ensuring that the model output conforms to the structure defined by the developer through JSON schema can make the model more reliable and secure.

The latest version of the model "GPT-4o-2024-08-06" scored in the JSON mode evaluation100%。

In comparison, the GPT-4-0613 version scored less than 40%.

For code editing, the latest version scores the same as the previous version on benchmarks, but now costs half the price.

Others have discovered that this new version is OpenAI'sThe cheapestModel.

API supports structured output

At last year's OpenAI DevDay, the official introduced JSON mode for GPT-4 to help developers build applications.

Although JSON Schema can improve model reliability, it does not guarantee that the model generated results conform to a specific schema.

Therefore, in this update, OpenAI introduced structured output in the API to ensure that the content generated by the model exactly matches the JSON schema provided by the developer.

There are two main forms:

1. Function call: Call structured output by setting strict:true in the function definition.

This feature is available for all model types, including gpt-4-0613 and gpt-3.5-turbo-0613. When structured output is enabled, the model output will match the provided tool definition.

2. New options for response_format parameter.

Developers can provide a JSON schema via json_schema (a new option for the response_format parameter).

This is useful when the model is not a calling tool, but responds to the user in a structured way. This feature is available for 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 has updated its Python and Node SDKs to support native structured output capabilities.

In addition, security output is emphasized. The new model will follow existing security rules and introduces a refusal string value to help developers identify situations where the model rejects unsafe requests.

Structured output can also be used for a variety of 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, enter the price as $2.5/1 milliontoken, the output price is 10 USD/1 million tokens.

Since the new model is more economical in input and output, the price is reduced after rounding off.

It was found that the output of this new model was less lengthy and more cost-effective than the previous version.

Reference Links:
https://openai.com/index/introducing-structured-outputs-in-the-api/