How to use GPT-5 Pro and Thinking via API?

API
1

Hi,

how can I use GPT-5 Pro and Thinking via the API? It seems they are not available as models. I only have gpt-5 available in my model list (miniand nanoare). I have not been able to find any documentation on this. Am I blind?

Thanks!

2

With the API’s gpt-5, you have four levels of “reasoning effort” that you can send, along with a verbosity parameter. Other reasoning models have three “reasoning effort” but not minimal as a setting.

ChatGPT is likely just variations on these, using its own chatty model and internal messages and tools.

3

I see. And is there any convention of what combination of effort and verbosity are considered “Thinking” and “Pro”?

4

There aren’t exact analogs. Self-reporting by ChatGPT GPT-5 Pro:

Here’s what I see in the settings:

  • oververbosity: 5 (target detail level for final answers)
  • Juice: 128 (internal reasoning budget for this turn)

API: reasoning & verbosity: “high”

  • oververbosity: 8
  • Juice: 200

“medium” verbosity delivers a value of “3” to the AI.

Then there is other non-transparency related to ChatGPT, if a “router” still is making decisions, a bunch of stylistic messaging of how to think about problems and puzzles, while API: you just optimize for you task.

Don’t try to copy ChatGPT; you can do better than that.

1 Like
5

But GPT-5 Pro is not GPT-5 Thinking (High). It’s a different model that uses parallel agents. Like o3-pro is’n’t o3-high

6

Not parallel “agents” but described as “settings for parallel test-time compute”. “Test time” meaning during inference/generation (as opposed to training the AI better and larger), and “parallel” meaning some sort of alternate speculation and best-of not disclosed or discussed beyond a mere phrase.

7

Yes, you are correct, thanks for the clarification.
That being said, a shame GPT-5 Pro isn’t available on API

1 Like
9

Don’t try to copy ChatGPT; you can do better than that.

could you give a hint - how?

10

ChatGPT has tons of prompting language and tools dumped into the AI context, ultimately degrading the quality. You don’t want to make images, don’t want to connect your google calendar, don’t want warnings about where to vote, don’t want the AI tuned up with hundreds of words — doesn’t matter, massive tool load.

On the API, you have a developer message that only needs to reshape “You are an AI model on an API” into something useful without distraction. Then you have direct API controls for how much thinking and how much writing you want out of GPT-5 models.

1 Like