Your OpenClaw agent stopped answering? Check the model before the server
A lapsed ChatGPT subscription silently gates the model your agent is set to: every turn fails while the machine looks healthy. The one-command escape, and the triage order that finds it fast.
When an agent that worked yesterday goes silent and the machine looks healthy (container up, messages delivering, nothing restarted), the first suspect is not the server. It is the model call. The commonest version: the ChatGPT subscription the agent signs in with lapsed or was downgraded, the account dropped to a tier that no longer includes the model the agent is set to, and every turn now dies with “model is not supported when using Codex with a ChatGPT account” while everything around it reports fine. The fix is one chat command, no restart: /model with the exact id of a model your account can still reach.
The failure, as we watched it happen
We run OpenClaw agents for a living, and this one played out on one of our own machines in late August 2026. A ChatGPT subscription connected to the agent quietly ended: no email anyone read, no error at the moment it happened. The account reverted to the free tier. The agent’s model stayed set to a pay-tier model. From then on, every single turn failed with an HTTP 400 from the provider, the “not supported” error above, while every layer an operator normally checks was green: the container was up, the Gateway was healthy, Telegram was delivering messages in, and the agent was receiving them. The only broken thing was the one call that produces an answer.
From the owner’s side of the chat, that is indistinguishable from a dead server, which is exactly why it is worth knowing about. You cannot ask a dead agent what is wrong; the diagnostic that works when the agent answers is unavailable in precisely the failure where you need it.
Why a lapsed subscription does this
An agent that signs in with a ChatGPT subscription inherits that subscription’s tier gates, permanently and live. Which models the connection serves is decided by the account’s current plan on every call, not by whatever was true when you connected: the plan tiers differ in exactly which models they include, so a plan change is a model-list change. And OpenClaw, by design, does not paper over it: current docs are explicit that if a configured model is unavailable to the account, you select an included one explicitly: “OpenClaw does not silently downgrade.” That is the right call for an agent (silent downgrades would mean silently different quality), but it means the failure mode is a hard stop, not a degradation.
The escape command
The lever that works is the in-chat model switch, because it needs no working turn to apply: it lands on the next message. The detail that matters is the full model id, provider prefix included, because the prefix decides which credential the call rides:
/model openai/gpt-5.6-sol # routes via the ChatGPT subscription
/model openrouter/openai/gpt-5.6-luna # routes via an OpenRouter API keyIn the incident above, switching to another openai/ model changed nothing: the route was still the dead subscription. Switching to the same family through the machine’s working API key brought the agent back on the next message, no restart, and the conversation resumed where it had stopped. If your setup has only the subscription and no second credential, the fix is a model the free tier still serves, or reviving the subscription. The subscription guide covers what each state serves.
The triage order for “my agent stopped answering”
1. Is the message arriving? If your chat shows delivery (Telegram ticks, typing indicators ever firing), transport is fine and the problem is deeper in.
2. Is the machine up? One look: docker compose ps (or your service manager) for the Gateway’s state. Up and healthy plus silence points onward.
3. Then suspect the model call, and read the logs. docker compose logs --tail 100 on the gateway service (or your install’s log file) will show the per-turn failure and its error text verbatim. A 400 naming the model, a quota error, an auth error. Each names its own fix, and all of them look identical from the chat window: silence.
Subscription tier gates are one member of a family. A spent API-key balance, a revoked key, a provider outage on one model: all produce a healthy machine whose every turn dies at the same door, and the same /model escape route applies to each while you fix the credential at leisure.
Know the way out before the door closes
The practical takeaway costs nothing today: while your agent is answering, learn what /model reports, note which second route your install has (another provider key, another model family), and remember that the command works when nothing else does. An agent owner who knows the escape command diagnoses this failure in one minute from a phone. One who does not is reading server logs for an infrastructure problem that does not exist.