[Feature]: Remember reasoning level for each model

[kiikoh]

Before submitting

• I searched existing issues and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/server

Problem or use case

With the release of GPT 5.5 I find myself switching back and forth between 5.4 depending on the task. It is very easy to kill your limits if you accidentally use 5.5 on medium/high reasoning. This is bound to happen if you keep switching back and forth

Proposed solution

Remember the last used reasoning on a per model basis

Why this matters

Prevents using too many tokens
Allows users to use different models in different ways

Smallest useful scope

T3 Code remembers the last used reasoning level for each model

Alternatives considered

No response

Risks or tradeoffs

What does this look like for existing users, how to backfill, etc? Users would not be able to set a default reasoning level globally... what happens when a new model is added? Should using the same model in different providers share the same stored reasoning level?

Examples or references

No response

Contribution

• I would be open to helping implement this.

[LPFchan]

i second this! having to move between gpt-5.6-sol medium and k3-256k high, i have switch reasoning effort/level every single time which is annoying

[id0Sch]

Same pain, but for service tier: I want Luna/Sol on Fast and Astra on Standard, and switching models currently carries or resets the wrong tier. Would love per-model memory for options generally (not only reasoning), or at least for serviceTier. @juliusmarminge wdyt?