GuidesRunning & maintaining

Codex: "model is not supported when using Codex with a ChatGPT account"

OpenAI's server refusing the model Codex sent: usually a retired name pinned in config.toml or a script, else a lapsed plan or an old Codex.

November 15, 2026The Everpod team
The short answer

“The ‘<model>’ model is not supported when using Codex with a ChatGPT account” is OpenAI’s server refusing the model name your Codex sent, for the account it is signed in with. It reads like a billing problem, but the commonest cause is a model name pinned in ~/.codex/config.toml, a profile or a script’s -m that OpenAI has since retired for ChatGPT sign-in: gpt-5.5 went on October 14, 2026, gpt-5.4 and gpt-5.4-mini on August 31. The other two causes are a plan that lapsed or changed, and a Codex too old for a new model. Find where the name comes from, replace it with one the account offers today (as of mid-November 2026, gpt-6.1-sol on paid plans), or delete the pin and let Codex choose.

What the message means

The sentence is not Codex’s own error text. It comes back inside the server’s answer, which is why reports on Codex’s tracker show it wrapped in JSON, such as {"detail":"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."} in one March report. The server is saying one thing only: for this ChatGPT sign-in, on this client, that model name is not served. It does not say whether the name was retired, never existed, or belongs to a plan you no longer have, and that ambiguity is what costs people the evening. The reporter of issue 43751 put it plainly after finding the cause: the message “reads as a plan or entitlement problem”, when it was “model rot on my side, not an entitlement problem.”

It only happens with a ChatGPT sign-in. OpenAI’s Codex models page says that with an API key, “model availability follows your API organization and project”, and that the retirement of gpt-5.5 from ChatGPT sign-in left “Codex sessions authenticated with an API key” unaffected. codex login status prints which of the two this machine is using.

Find the model name your run sent

A non-interactive run prints the model it is using before it starts. From the 0.152.0 report, a bare codex exec on an account whose config still named a retired model:

$ codex exec "Reply with exactly: OK"
model: gpt-5.1
warning: Model metadata for `gpt-5.1` not found. Defaulting to fallback metadata
ERROR: {"status":400,"error":{"message":"The 'gpt-5.1' model is not supported when using Codex with a ChatGPT account."}}
Copied

The warning above the error is the better clue: Codex did not recognise the name either, which points at a retired or mistyped model rather than at your plan. In an interactive session, /status shows the active model.

Then find where the name is set. Codex takes the model from the first of these that names one, highest first, per its configuration docs:

Inside a session, /debug-config prints the layers Codex loaded. From a shell, this shows the model lines in your own files:

grep -n '^model' ~/.codex/config.toml ~/.codex/*.config.toml
grep -rn -e '--model' -e ' -m ' <your-scripts-folder>
Copied

If no layer names a model, Codex uses its recommended one, so deleting a stale pin is a fix in itself. The reporter of issue 43751 did exactly that, and Codex then picked a current model on its own.

The three causes, and how to tell them apart

The model was retired for ChatGPT sign-in. This is the usual one for a setup that worked for months. OpenAI’s models page, as of November 15, 2026, lists these as gone from Codex with ChatGPT sign-in:

The current Codex models on that page are Astra (gpt-6-astra), 6.1 Sol (gpt-6.1-sol), 6.1 Terra (gpt-6.1-terra) and 6 Luna (gpt-6-luna). Check the page before you pin one: it is the list that moves.

The plan no longer covers the model. A subscription that lapsed or was cancelled drops the account to a plan with a much shorter model list, and the model that worked yesterday is refused today. On the March thread a user wrote, “Turned out my ChatGPT Plus subscription had expired — the error disappeared immediately after I renewed it.” Check the plan in ChatGPT’s settings for the account Codex is signed in to, which is not always the one you think on a server signed in months ago. Two plan cases are not yours to change: on Enterprise and Edu “an administrator must enable Luna first”, and a ChatGPT workspace can deliver configuration defaults to Codex, so a refused model you never set may be your admin’s to update. What each plan includes is in which ChatGPT plans include Codex.

The client is older than the model. Codex asks for its model list with its own version number, so a new model can be refused to an old client on a plan that includes it. OpenAI states one such floor: “GPT-6.1 Terra requires Codex CLI 0.159.0 or later.” If the name is current and your plan includes it, run codex --version, update Codex the way you installed it, and try again; the latest release on November 15, 2026 was 0.169.0.

A fourth pattern is in the older threads: the refusal arriving for a current model on a paid plan and clearing by itself, in ten minutes for the March reporter and within days for a gpt-5.5 report in April. If none of the three causes fits, that is the one left, and only OpenAI can see it.

See what the account can use

Interactively, /model lets you choose from the models Codex offers this sign-in, which avoids typing a dead name. For a script, codex debug models (marked experimental) prints the model catalog Codex sees as JSON; look for the slug values. Codex also keeps the list it was last given in models_cache.json in its home folder, normally ~/.codex, and asks again after five minutes. To check a name before you commit it to a config file or a job:

codex exec -m gpt-6.1-sol "Reply with exactly: OK"
Copied

On a server, and in scheduled jobs

A server is where a pinned model sits longest without anyone looking. A cron job or a systemd timer written in the spring with -m gpt-5.4 on a ChatGPT sign-in fails on every run from August 31, and codex exec exits with status 1 when a turn fails, so a job that does not check the status records nothing at all. Three things differ from a laptop:

For a job that should keep working across model changes, either leave the model unset so Codex picks its recommended one, or give the job an API key, whose models follow your API organization instead of your plan and are billed at API rates rather than drawn from the plan. How an unattended run reports a refused turn, and how to wrap one, is in the guide to the capacity error; the setup of Codex on a VPS is in Codex CLI on a Linux server. A plan’s limit is a different refusal with its own wording, covered in “You’ve hit your usage limit”.

OpenClaw and other agents on a ChatGPT sign-in

Agents that sign in with a ChatGPT account through Codex’s route, OpenClaw among them, get the same refusal, because the server is answering the same question. The causes are the same three, with one twist on the third: the agent announces the version of the Codex it bundles, not the one you installed. On OpenClaw 2026.9.7 and 2026.9.8, which bundled Codex 0.158.0, users had gpt-6.1-sol refused with this message; one concluded in the issue, “So the gate is the client version advertised to the backend, not the account.” Updating the agent is the fix there. In an agent, the message often never reaches you: the agent simply stops answering, which the guide to an OpenClaw agent that stops responding walks through, and which model to set instead is in the best model for OpenClaw.

Run Codex on an always-on developer pod.

A developer pod is a permanent cloud computer (VPS), set up and secured for you, with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed. Run more agents at once than your laptop can hold, and keep them working when it’s closed. From $24 a month, built in about ten minutes.