GPT-5.6 Sol Pro in Chat mode reverting to Instant-level routing or reasoning budget
What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.715.31925
What subscription do you have?
Pro 20x
What platform is your computer?
Darwin 25.5.0 arm64 arm
What issue are you seeing?
When I manually select GPT-5.6 Sol Pro in Chat mode within the unified ChatGPT/Codex app, prompts that would ordinarily trigger a substantial Pro reasoning run now complete in approximately 5 seconds.
The resulting answers are materially shallower, less careful, and lower-quality than previous GPT-5.6 Sol Pro responses. They appear behaviorally similar to responses from an Instant model or from a model running with a substantially reduced reasoning budget.
There is no visible error message. The interface continues to display GPT-5.6 Sol Pro as the selected model, but the latency and output quality do not appear consistent with the selected mode.
This behavior is reproducible in both:
the unified desktop app; and
the unified ChatGPT web interface.
It has also been independently reproduced on another ChatGPT Pro 5x account.
I cannot determine from the client whether the wrong model is being selected at inference time or whether GPT-5.6 Sol Pro is receiving an unexpectedly low effective reasoning budget. The observable symptoms are consistent with either:
silent routing or fallback to an Instant model;
incorrect model configuration or metadata;
an unexpectedly reduced reasoning-effort setting;
an account-level or rollout feature-flag problem; or
a regression associated with the recent unification of the ChatGPT and Codex interfaces.
What steps can reproduce the bug?
Selecting "Pro" reasoning effort where the model is GPT-5.6 Sol
What is the expected behavior?
_No response_
Additional information
_No response_
7 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Example:
no-active-thread-019f75ff-f3d1-7263-8546-91449088f8d6
same here.
This shit is real
same on my side
I can reproduce this on ChatGPT web with a ChatGPT Plus account.
I verified that this was an actual routing downgrade, not merely the assistant incorrectly self-identifying its model. There was no visible fallback, capacity, eligibility, or rate-limit warning.
Expected behavior: the request should run on the explicitly selected GPT-5.6 Sol with High reasoning. If that route is unavailable, ChatGPT should show a clear error or ask for confirmation before using a lower-tier model.
Actual behavior: a lower-tier model was served silently while the selector still displayed GPT-5.6 Sol / High.
Adding a concrete behavioral example from a ChatGPT Plus account.
who is jansci“Jancsi” had recently been used in another conversation on the same account to mean János/John von Neumann, whose familiar Hungarian name was Jancsi. Chat answered:
Cross-conversation recall is not guaranteed, so selecting another plausible Hungarian referent is not by itself diagnostic. The concerning behavior is that the model invented a detailed shared history to justify its guess, converted language from the current exchange into something that supposedly happened “yesterday,” and then claimed to have recovered the referent without pretending.
Work mode on the same account, with the relevant context available, immediately resolved Jancsi as von Neumann and identified the fabrication. This example cannot establish the backend route by itself, but it shows the same observable quality discontinuity: weak contextual discrimination and confident confabulation in Chat while the UI continues to show Sol. If Chat silently falls back or runs with a substantially reduced reasoning budget, the interface should disclose that change.