The 'gpt-5.6-sol' model is not supported when using Codex with a ChatGPT account.

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #33023
• #33146
• #33214

Powered by Codex Action

What a strange error message. Am I supposed to assume it's supported when using a non-ChatGPT account?

Sol disappeared from latest update

what do you mean by disappeared? disappeared for all users?

<img width="546" height="177" alt="Image" src="https://github.com/user-attachments/assets/b47faf74-06c6-4601-b181-d529725e8605" />

Pro subscriber.

it was set on custom after restarting it

<img width="428" height="77" alt="Image" src="https://github.com/user-attachments/assets/d297acd1-f890-4d06-96e2-b46adc9ee9b6" />

<img width="772" height="549" alt="Image" src="https://github.com/user-attachments/assets/52526ee9-8bdc-49fd-9169-ee210ce8e3f0" />

Turns out it was my sub that just didn't auto renew.

Additional repro, and one detail I have not seen reported in this thread: an unrecognised model id does not error — it is silently served by a different model.

Environment

• codex-cli 0.146.0
• auth mode: chatgpt-account (codex login device-code flow; codex login status → "Logged in using ChatGPT")
• plan: ChatGPT Pro
• Linux x86_64

Observed

$ codex exec -m gpt-5.6      "..."
ERROR: {"type":"error","status":400,"error":{"type":"invalid_request_error",
 "message":"The 'gpt-5.6' model is not supported when using Codex with a ChatGPT account."}}

$ codex exec -m sol          "..."   # same 400

$ codex exec -m gpt-5.6-sol  "..."   # NO error
gpt-5.4-mini                          # <- self-reported model in the response

$ codex exec -m gpt-5.6-terra "..."  # works
$ codex exec -m gpt-5.6-luna  "..."  # works
$ codex exec                  "..."  # default → gpt-5.4

The part worth prioritising: gpt-5.6 and sol fail loudly with a 400, which is correct and actionable. gpt-5.6-sol — the documented id — fails silently and appears to be served by gpt-5.4-mini, with no warning, no metadata notice, and exit code 0. A user who pins gpt-5.6-sol in ~/.codex/config.toml has no signal that they are running a substantially smaller model.

We only caught it because we require a firing transcript before trusting any model pin. Without that step we would have run a cross-model review seat on 5.4-mini while believing it was Sol.

Caveat on the evidence: the CLI does not appear to expose the served model in --json event output, and model self-reports are unreliable (terra self-reported "gpt-5.6", luna "gpt-5"). So gpt-5.4-mini is the model's own claim, not an authoritative identifier. Surfacing the served model id in the JSONL event stream would make this class of substitution detectable by users regardless of how the routing bug is resolved.

A 400 is much better than a silent downgrade.