Add "minimal" reasoning effort for GPT 5 models

[hf-fdsilva]

What feature would you like to see?

Based on a few of the OpenAI specific documentation, the GPT-5 models support a new "minimal" reasoning effort that works much better for coding tasks. Unfortunately, this current reasoning is not exposed as an option in Codex CLI.

Trying the following run command:
codex --ask-for-approval never --sandbox danger-full-access -c model_reasoning_effort="minimal"
Results with the following error:

Error loading configuration: unknown variant `minimal`, expected one of `low`, `medium`, `high`, `none`
in `model_reasoning_effort`

Regarding the minimal reasoning effort functionality, please refer to the following articles:
"GPT‑5 with minimal reasoning is a different model than the non-reasoning model in ChatGPT, and is better tuned for developers." (Source)

"The new minimal setting produces very few reasoning tokens for cases where you need the fastest possible time-to-first-token. We often see better performance when the model can produce a few tokens when needed versus none. The default is medium.

The minimal setting performs especially well in coding and instruction following scenarios, adhering closely to given directions. However, it may require prompting to act more proactively. To improve the model's reasoning quality, even at minimal effort, encourage it to “think” or outline its steps before answering." (Source)

Are you interested in implementing this feature?

No response

Additional information

No response

[aehlke]

The string is "low" as it says in the error

edit: I was wrong

[hf-fdsilva]

Is "low" properly setting the reasoning effort to "minimal" for the GPT-5 specific models? The OpenAI card mentions that this new model specifically requires the effort to be set to minimal :/

See below:

The reasoning.effort parameter controls how many reasoning tokens the model generates before producing a response. Earlier reasoning models like o3 supported only low, medium, and high: low favored speed and fewer tokens, while high favored more thorough reasoning.

The new minimal setting produces very few reasoning tokens for cases where you need the fastest possible time-to-first-token. We often see better performance when the model can produce a few tokens when needed versus none. The default is medium.

The minimal setting performs especially well in coding and instruction following scenarios, adhering closely to given directions. However, it may require prompting to act more proactively. To improve the model's reasoning quality, even at minimal effort, encourage it to “think” or outline its steps before answering.

[mohammad-malik]

On a related note, its unclear if setting it to High counts subtracts from GPT-5 Thinking caps (3000 per week for Plus) or GPT-5 caps (160 per 3 hours for Plus).

Hope these are both addressed soon!

[aehlke]

@mohammad-malik the rate limits in codex are independent of the chatgpt app/website. and codex only uses the -thinking models.

[mohammad-malik]

@mohammad-malik the rate limits in codex are independent of the chatgpt app/website. and codex only uses the -thinking models.

@aehlke Thats news to me? I thought they made it so Codex-CLI, when signed in with ChatGPT, now counts towards the ChatGPT limits? If not, what are the limits? How do things work here then? I know its not the 5$ API credits anymore.

[aehlke]

@mohammad-malik the rate limits in codex are independent of the chatgpt app/website. and codex only uses the -thinking models.

@aehlke Thats news to me? I thought they made it so Codex-CLI, when signed in with ChatGPT, now counts towards the ChatGPT limits? If not, what are the limits? How do things work here then? I know its not the 5$ API credits anymore.

check https://x.com/embirico (once online) for these updates

no codex limits published yet as they are still in flux

[techjoec]

@seratch - i couldnt find where to really comment here on your commit/PR since it's done and merged now but:

I just wanted to say a huge THANK YOU THANK YOU THANK YOU for this. I was fighting the long slow replies for MANY MANY hours and happened to randomly see your PR. I compiled almost immediately and was using it all night. Huge quality of life thing! You rock!