"This model is not available in your region" on a server
OpenRouter and the labs refuse models by the country your server's traffic comes from, not yours. How to check that country and what clears it.
“This model is not available in your region.” is OpenRouter refusing a model whose maker does not serve the country your request comes from. Anthropic, OpenAI and Google each publish a list of the countries they serve, and the check that raises this error looks at where the request leaves from: for an agent on a server, that is the data centre’s country, so a server outside the lists is refused even when you are inside them. Servers in Hong Kong recur in the reports, and none of the three lists Hong Kong, mainland China or Russia as of November 2026. Run curl https://ipinfo.io/country on the server to see where its traffic appears to be. If you are somewhere the lab serves, the fix is a server in a country on its list, or a model whose maker serves the server’s country. The lists also cover where the service is used and by whom, so a server elsewhere is not a way to use a lab from a country it does not serve. Your key is fine.
The same refusal in each provider’s words
Which wording you see depends on whom the agent calls. Through OpenRouter, the body is short:
{"error":{"message":"This model is not available in your region.","code":403}}CopiedSome reports show metadata with it, a routing_funnel whose failed_routing_step is “Gate Endpoints with Geo Restrictions”, as in a Hermes report and an oh-my-pi one. The step’s name says what happened: OpenRouter took the model’s endpoints and found none it could use from your location. OpenRouter’s routing documentation classes 403 and 451 as “geographic or legal blocks”.
Calling a lab directly, the refusal reads differently:
- OpenAI: HTTP 403 with code
unsupported_country_region_territory, typerequest_forbiddenand the message “Country, region, or territory not supported”. OpenAI’s error codes page gives the cause as “You are accessing the API from an unsupported country, region, or territory.” - Google (Gemini API): HTTP 400, status
FAILED_PRECONDITION, “User location is not supported for the API use.” - Anthropic: its errors page documents no region message. Users report a 403 with type
forbiddenand the message “Request not allowed”, which one thread traced to the region. Claude Code on Amazon Bedrock has printed “Access to Anthropic models is not allowed from unsupported countries, regions, or territories”, with a 400, from a Bedrock region in mainland China that worked from Seoul.
Other routers pass on their own version: Hermes users on one gateway have seen “This model is not available in your country.”
Why the server’s country is what the check sees
The check that produces this error looks at the place the request comes from, and an agent’s requests come from the machine it runs on. Google says it outright for Colab on its available regions page: “Region restrictions are applied based on the region that the Colab instance is in, not the region that the user is in.” The same holds for a server you rent. An owner in France with an agent on a Hong Kong server is, to the lab, a caller from Hong Kong.
Going through OpenRouter does not change whose list applies. Anthropic’s supported countries policy covers “Claude models offered through cloud platforms and authorized resellers” as well as its own API and apps, so a Claude model on OpenRouter is refused wherever Anthropic’s API would be. That is why the refusal lands on some models and not others from the same server with the same key: one team on a Hong Kong server reported Claude, Gemini and GPT models refused through OpenRouter in March 2026, while DeepSeek, Qwen, MiniMax, GLM and Kimi models answered.
The lists, as published by each lab:
- Anthropic’s supported regions (territories of a listed country are covered too, so Puerto Rico counts as the United States)
- OpenAI’s supported countries and territories
- Google’s available regions for the Gemini API
The three lists differ, so a server in one country can reach one lab’s models and not another’s. The usual hosting countries (Germany, Finland, the United States, Singapore, Japan, the United Kingdom) are on all three as of November 2026.
Which country your server’s traffic comes from
Run this on the server itself, not on your laptop:
curl https://ipinfo.io/countryCopiedIt prints a two-letter country code such as DE or HK. Plain curl ipinfo.io, the check Google’s page gives for Colab, prints more about the same address.
It is a check, not the lab’s verdict. Each provider uses its own idea of where an address is, and the databases disagree at the edges. Gemini API users have reported the location error from a server in Toronto and from one in the Netherlands, both on Google’s list, and a German server owner on the Hong Kong thread above found his host’s addresses had once been registered to Iran. If the check shows a listed country and the refusal persists, that is a matter for the provider’s support, with the server’s address and the full error.
What clears it
- A server in a country on the maker’s list. If the model you want is Claude, GPT or Gemini, the agent has to run somewhere Anthropic, OpenAI or Google serves. Moving it means a new server in another region and the agent’s home directory copied across. Between the usual hosting regions, a short model call differs by a few tenths of a second, as the region timing guide measured.
- A model whose maker serves the server’s country. On OpenRouter, switch the agent to a model from a lab that serves that location. Check it from the server with one real request before relying on it, since the refusal comes per model.
- A cloud platform’s region, when the agent uses one. On Amazon Bedrock, the Claude Code report above was refused in a mainland China region and answered from an instance in Seoul, so the platform region the agent calls is worth checking too.
Both OpenAI and Anthropic say their services are for the places on their lists, and OpenAI’s page adds that access from elsewhere “may result in your account being blocked or suspended.” Those terms cover where you use the service from, not only where the server is: moving the agent clears the error for someone the lab serves, and is no way round a list for someone it does not.
It is not your key
The refusal is a 403, and some agents treat every 403 as a bad credential. In Hermes, as of November 2026, an unsupported_country_region_territory refusal is shown as a rejected API key with a pointer to hermes setup; the report and its fix were still open. Rotating the key changes nothing. To settle it for OpenRouter, query the key’s own details as the OpenRouter key guide shows: a key that answers there is working, and the region is the problem.
Three OpenRouter messages look close and mean something else:
- A 403 about the domain the request arrived on. A workspace guardrail can limit a key to a regional domain such as
eu.openrouter.ai, and a request through any other domain is rejected. That is about the address the agent calls, not where the server is, and is changed in the guardrail or the agent’s base URL. - “No endpoints found”. No provider for the model takes what the request needs, such as tools or images. The fixes are in the no-endpoints guide.
- “Provider returned error” with a 403 inside. The provider itself refused, and its own message is in the error’s metadata, as the provider error guide explains; if that inner message names a country or region, it is the same problem as this page.