How to Make OpenAI API and Web GUI Perform Identically, Especially in the "Thinking Process"

[xhy20230601]

I am using OpenAI API and the ChatGPT web GUI for image recognition tasks, and I have noticed a difference in their performance, particularly in the “thinking process.”

When using the web GUI, the model seems to generate code for recognizing specific items within the image and even performs a search for double-checking. On the other hand, the API only provides pure analysis and gives the final conclusion. Specifically, the GUI appears to handle finer details of the image and performs step-by-step reasoning, while the API’s output is more concise and limited to a final analysis.

I would like to make the API perform exactly like the web GUI, especially in terms of the thinking process (e.g., recognizing individual items and performing a double-check). Is there a way to adjust API settings or apply any method to make the performance of both the API and the GUI consistent?

Any advice or suggestions would be greatly appreciated!

[aprendendo.next]

@xhy20230601

When using the web GUI, the model seems to generate code for recognizing specific items

That is not possible at this moment because the completions/responses API doesn’t support code interpreter, and the assistants API doesn’t support the newer models used in ChatGPT.

Basically, ChatGPT is a product that although it uses the same models as in the API, internally have a different toolset of extensions that makes it unique. The same behaviour can only be replicated for certain use cases, but not all of them.

[xhy20230601]

Thanks for your reply.

So that means openai are using some specific extension tools for only web version(GUI), but these tools are not avalible for API users?

The ability of model is the same, how should I make API responds expose more details of thinking process? Could it works if I design a strong prompt?

[aprendendo.next]

@xhy20230601

So that means openai are using some specific extension tools for only web version(GUI), but these tools are not avalible for API users?

Yes. For example, web search is not available on certain models in the API.

@xhy20230601

how should I make API responds expose more details of thinking process? Could it works if I design a strong prompt?

You don’t control these, only the parameters.

You can read more about it in the docs:

• https://platform.openai.com/docs/guides/reasoning?api-mode=chat
• https://platform.openai.com/docs/api-reference/responses/create#responses-create-reasoning

[_j]

@xhy20230601

The ability of model is the same,

ChatGPT and the API run different models, even of “gpt-4o”. Developers didn’t get AI that turned into a user praise-machine when ChatGPT did.

Huge difference between ChatGPT assistant and API assistant

It is not an issue. It is expected. ChatGPT is a realized product, with its own model, its own internal prompting, and its own non-replicable tools. The API is a development platform. You won’t have your own product invoking a special tool injection to then tell people where to vote, for example, but ChatGPT does. You can’t receive custom “tool” output (only functions) and you haven’t written tons of anti-infringement text for the AI, but that’s what DALL-E in ChatGPT does. Files work diffe…

[aprendendo.next]

@_j

ChatGPT and the API run different models, even of “gpt-4o”. Developers didn’t get AI that turned into a user praise-machine when ChatGPT did.

Wasn’t that available even for the model chatgpt-4o-latest ? (Which supposedly is different thant gpt-4o)

I don’t use that though, as it is sort of unpredictable…

[_j]

Yes, but then you move to inability to verify what is offered is actually what is being run. You are blocked from using any functions (just one type of tool) - and ChatGPT doesn’t have just a functions tool for passing yours, it is loaded with unplaceable tool recipients and their internal instructions with any API model.