OpenCode "Upstream request failed: Endpoint is unavailable"
OpenCode's Zen and Go gateway failing, not your install: check the status page, try another model, report it. Plus "Model unavailable: opencode-go/…".
“Upstream request failed: Endpoint is unavailable.” is an answer from OpenCode’s hosted gateway, the service behind the OpenCode Console (Zen) and OpenCode Go models, not a fault in your install or your config. First check status.opencode.ai: on the main issue, an OpenCode maintainer traced a day of these errors to a degraded period there, “a service outage rather than a bug in the app”. If the status page is clear and only one model fails, switch to another with /models and report the model and the time to OpenCode’s support at help@anoma.ly. Reinstalling or updating OpenCode does not change it.
Where the message comes from
OpenCode sends requests for its own models, the opencode/… and opencode-go/… ones, to OpenCode’s gateway, and prints what the gateway answers. The body the gateway returns, as reporters have captured it with curl:
{"type":"error","error":{"type":"api_error","message":"Upstream request failed: Endpoint is unavailable."}}CopiedOpenCode’s triage bot says the error “comes from the OpenCode Zen/Go gateway, and several different problems produce it,” and on later reports, “not from the OpenCode app.” Reports show it with an HTTP 403, a 422 or no status at all, and from Claude Code and plain curl pointed at the gateway as well as from OpenCode, which is why nothing on your machine changes it.
The three patterns in the reports
- Everything from OpenCode fails at once. On October 22, 2026, free Console models and Go models failed together while, in one reporter’s words, other providers such as OpenRouter kept working. That was the outage the maintainer matched to the status page. The maintainer’s advice since: “If it happens again now, check https://status.opencode.ai first,” and if it is clear, open a new issue with the model and the time.
- One model fails, the rest work. In November 2026, Go subscribers reported
gemini-4-argonand laterclaude-haiku-5-5returning the message on every request for hours, while every other Go model on the same key returned normally and the model list still showed the failing one. Neither thread records a cause or a fix; both were closed with a pointer to email support, because “Support for OpenCode Go and the Console isn’t handled in GitHub issues.” - Only some turns fail. One report on a free model had plain chat working and every turn after a tool call failing with a 422. The bot asked for the failure times and the Console request IDs, by email.
The status page tracks the Console, the gateway, Go, pay-as-you-go and the free models, and for each plan only its six most-used models. A single model failing while the page shows green is therefore not a contradiction, and is the case to report.
What to check, in order
- The status page. A degraded Go, pay-as-you-go or Free inference line means waiting is the fix.
- Another model. Pick one from
/modelsin the TUI. If others from the same plan answer, the problem is that one model; if none from OpenCode answer, it is the service. - A model from another provider. A provider connected with your own key, through
/connect, does not go through OpenCode’s gateway, so an outage of OpenCode’s does not reach it; choose it in/modelsto carry on. Choosing models in OpenCode covers connecting one. - A report. Email help@anoma.ly with the model, the time in UTC and, if you can find them in the Console’s request logs, the request IDs. Issues on GitHub about the hosted service are closed and redirected there.
A subscription that has reached its limit, or a bad key, gives different messages; this one does not mean either. The key’s own error is covered in OpenCode’s “Invalid credential” error, and how Go’s limits and the Console balance work in OpenCode’s pricing.
“Model unavailable: opencode-go/…”
A message that looks related comes from OpenCode itself, before any request leaves your machine:
Model unavailable: opencode-go/deepseek-v4.1-flashCopiedOpenCode’s providers docs give its meaning: “The provider is inactive, the model ID is absent or disabled, or dynamic discovery no longer returns it.” In practice that is one of three things. The provider is not connected (run opencode auth list to see). The model ID is not one the catalogue holds any more: Console models get retired and free ones are offered for a limited time, so a model named in opencode.json can outlive its listing. Or free models are hidden by OPENCODE_DISABLE_FREE_MODELS, in which case, per the models docs, “explicitly selecting one fails with Model unavailable.”
See what OpenCode currently offers with:
opencode modelsCopiedand choose from /models rather than typing an ID; the docs’ own advice is to “pick an entry from /models instead of guessing its provider or model ID.”
One open report, on OpenCode 2.3.0, describes a fourth case: after the machine lost its connection to opencode.ai, the background service kept a stale list and refused a free model with this message, an opencode service restart did not bring it back, and the list caught up on its own about four and a half hours later. OpenCode’s bot replied that the service fetches the Console’s model configuration at startup and checks it every minute after that; no maintainer has answered yet. If your machine was recently offline and the model is in opencode models elsewhere, that report is the one to follow, and OpenCode’s background service covers restarting and reading its logs.