Make reasoning-effort levels stable and transparent for subscription users

What variant of Codex are you using?

Codex App, CLI/TUI, and IDE extension with a ChatGPT subscription

What feature would you like to see?

Problem

Codex already lets users select reasoning levels such as high, xhigh, and max. The problem is that the effective reasoning budget behind those labels can apparently be changed server-side while the UI continues showing the same selection.

OpenAI recently acknowledged running experiments in which reasoning-effort values (“juice values”) were changed and later reverted:

https://x.com/thsottiaux/status/2076495156757577895

This means that GPT-5.6 Sol / xhigh can represent one level of reasoning today and a lower level tomorrow without any visible change in the model picker.

Calling this an “optimization” does not solve the transparency problem. If an optimization preserves quality, that is beneficial. If it reduces the effective reasoning available to the model, the same label should not remain unchanged.

For paid subscribers, this creates a poor consumer experience:

• model quality can change during an active subscription;
• the selected reasoning level no longer represents a predictable capability tier;
• coding results and evaluations become difficult to reproduce;
• users cannot distinguish a model regression from changes in their prompts or code;
• usage savings may come from an undisclosed reduction in reasoning quality.

Requested behavior

Please make reasoning-effort selections stable and transparent:

1. Keep manually selected reasoning levels semantically stable.
2. If their effective meaning changes, introduce a new profile revision or clearly notify users.
3. Add an Auto / Optimized option for users who accept dynamic server-side tuning.
4. Allow users to select a Stable / Pinned profile that is excluded from effort-changing experiments.
5. Show the effective reasoning level and profile revision in /status and session metadata.
6. Document material reasoning-profile changes in the Codex changelog.
7. Never silently map a manually selected level to a lower effective tier.

This does not require exposing private chain-of-thought or exact internal token budgets. A visible profile revision or another stable identifier would be sufficient.

API users can explicitly send a reasoning effort with each request. Subscription users should receive comparable control and should know when the effective behavior behind their selection changes.

An unchanged setting such as xhigh or max should continue representing a comparable quality tier. Materially changing its meaning without notice during an active paid subscription is not transparent or consumer-friendly.

Additional information

_No response_