Expose xhigh reasoning on supported GPT models
Now that GPT-5.2 supports an “xhigh” reasoning option (first introduced with the codex 5.1 models I believe), it would be useful to be able to select it, alongside the normal “low”, “medium” and “high”.
Many thanks :)
Also for Mini and Nano variants.
Log in to comment and vote
Comments6
Mark Florkowski
For anyone upvoting this, I would like to know more about what you think this would enable. In many (most) cases in the “chat” context, higher reasoning effort does not have notable gains, and just costs more.
Emerald Alligator
Not that long ago I wanted xhigh to review a plan for building a new app that I developed with another model. Instead of staying in t3chat I had to make a script whose sole purpose is to pass that one question to xhigh and get a response. Seems arbitrary to draw the line at exactly this one thing when we have all these other options with some being more efficient than others. Why even have the low, medium and high options then?
Mark Florkowski
The distinction isn't arbitrary, but rather based on the diminishing returns. In the "chat" context, the "High" setting is generally plenty (and often even overkill) to handle 99% of complex reasoning tasks efficiently. Many users look at the options and simple pick the highest one because “bigger number better”
That said, we may consider the option to allow these “enthusiast”/special cases behind some toggle/bringing your own key, if users want to accept the costs, without exposing it by default
Rose Octopus
Even though, you're right it is diminishing returns, it might prove to be of use in cases with complex math/physics problems, where less reasoning might fail.
Mark Florkowski
Possibly, that is why I want to hear about use cases
Teal Celestial
I use it when med/high gets something wrong, like a STEM problem for example.