Codex Desktop wrongly routes GPT-5.5 to Responses Lite, causing avoidable failures and wasted tokens

What issue are you seeing?

Hello OpenAI Support,

I want to report a Codex Desktop behavior that is unreasonable and significantly harms user experience.

Issue:
In Codex Desktop on Windows, GPT-5.5 is shown as an available model, but when selected it fails with:

"This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite"

GPT-5.4 works normally in the same environment.

Why this is a product issue:

1. The UI presents GPT-5.5 as available, which strongly implies it is usable.
2. After investigation, Codex Desktop regenerates local model metadata that marks GPT-5.5 with use_responses_lite: true, while GPT-5.4 is false.
3. This creates a misleading and self-contradictory experience:

• the model appears selectable
• the app routes it through a path it does not support
• the request then fails

1. This wastes user time, interrupts workflow, and can waste tokens on failed attempts.
2. This behavior is not aligned with normal product expectations. If GPT-5.5 is not currently supported on the Lite route for this client/version, it should either:

• not be shown as selectable, or
• be routed through a supported path automatically, or
• present a clear pre-selection warning before the user attempts to use it

What I already verified locally:

• Safe cache cleanup was completed successfully.
• auth.json, cap_sid, and other auth-related files were not modified.
• models_cache.json was deleted and regenerated by Codex Desktop.
• After regeneration, GPT-5.5 still appears with use_responses_lite: true.
• GPT-5.4 remains use_responses_lite: false.
• config.toml remains on wire_api = "responses".
• Therefore this does not appear to be just stale local cache; it looks like current capability/routing metadata being reissued to the client.

Requested action:
Please treat this as a product bug / routing bug in Codex Desktop.
At minimum, please prevent users from being offered a model that is known to fail on the currently assigned transport path.
This behavior causes unnecessary failures, wasted tokens, and a poor user experience.

Environment:

• Product: Codex Desktop official Windows app
• OS: Windows
• Date observed: 2026-07-09
• Error message:

"This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite"

If needed, I can provide:

• screenshots
• regenerated models_cache.json evidence
• exact timestamps
• local app/runtime version details

Please escalate this to the Codex Desktop / model routing team.

Thank you.

What steps can reproduce the bug?

you can scan the config

What is the expected behavior?

_No response_

Additional information

_No response_