model not found / model_not_found — one message, three unrelated causes

This error does not prove the model name is wrong. It equally often means your account isn't entitled to that model, or that the endpoint rewrites model names and yours isn't in its table. Trying different spellings fixes exactly one of the three. One request tells you which one you have: ask the endpoint what it actually serves.

What you're seeing

API Error: 404 {"type":"error","error":{"type":"not_found_error","message":"model: <model-id>"}}

{"error":{"message":"The model `<model-id>` does not exist or you do not have access to it.","type":"invalid_request_error","param":null,"code":"model_not_found"}}

404 model not found

{"type":"error","error":{"type":"not_found_error","message":"Not Found"}}

The second one is the honest one — it names both possibilities in the same sentence and declines to say which. Treat the other two as saying the same thing less clearly. The "or you do not have access to it" half is not filler: it points at 403 Forbidden territory wearing a 404's clothes.

The fourth is an impostor. It contains no model name, and that is the tell: a not_found_error with no model in it is usually the route being missing, not the model. See step 4.

Three causes, one message

Cause B is the one the wording hides. Semantically it's an authorization result — the correct answer would be a 403 with permission_error, which is a different page: 403 Forbidden. Many endpoints return 404 instead, deliberately, so that a probing request can't be used to enumerate which models exist. The side effect is that you spend an hour on spelling for a problem that has nothing to do with spelling.

Why the short name works and the full name doesn't

A compatible endpoint is under no obligation to serve the same catalog as the upstream provider, and most don't. What they publish instead is an alias table: short, undated tier names — haiku, sonnet, opus — each routing to whichever model that endpoint has selected for that tier. Some also substitute a default when a request names no model at all.

Two things follow, and both look like bugs when you don't expect them:

• The dated, fully-qualified ID you copied out of provider documentation may simply not be a key in that table. It is a perfectly valid model ID that this particular endpoint has no entry for. 404 on the full name, 200 on the short name, is the designed behaviour of an alias-mapped endpoint — not a fault, and not something a different spelling will get around.
• A short alias is not a stable identifier. What sonnet resolves to can change on the endpoint's side, with no change to your config. Convenient today; unreproducible when you come back to a result in three months and can't say which model produced it.

For the same reason this page does not reproduce anyone's alias table, and you shouldn't trust one you find written down elsewhere. What a tier name resolves to, and whether a request that names no model gets a substitute at all, are properties of the endpoint on the day you ask it. Which is why the next section starts with reading the catalog rather than guessing at names.

Check in this order

# the authoritative list — whatever this prints is what you are allowed to ask for
curl -s https://api.9coding.com/v1/models \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01" | jq -r '.data[].id'

curl -s https://api.9coding.com/v1/models \
  -H "x-api-key: sk-9c-xxxxxxxx" \
  -H "anthropic-version: 2023-06-01"

printf '%s' "$ANTHROPIC_MODEL" | od -c | tail -3

env | grep -i -E 'anthropic|model'

curl -sS -o /dev/null -w '%{http_code}\n' https://api.9coding.com/v1/models
curl -sS -o /dev/null -w '%{http_code}\n' https://api.9coding.com/v1/messages

echo "$ANTHROPIC_BASE_URL"   # exactly https://api.9coding.com — no trailing slash, no /v1

# <model-id>: copy an id from https://api.9coding.com/v1/models
curl -sS -w '\n%{http_code}\n' https://api.9coding.com/v1/messages \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{"model":"<model-id>","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}'

When it isn't your problem

• Your ID is in /v1/models and requests to it still 404. The catalog and the router disagree. Nothing on your machine can reconcile them.
• It worked yesterday and you changed nothing. Either a snapshot reached the end of its life upstream, or the endpoint's alias table moved under you. Both are events with a date; both should appear in a changelog.
• Other models on the same endpoint work fine. That rules out credentials, base URL, and network in one observation — you are looking at cause A, B, or C, never at a connectivity problem.
• A short name works and the full name doesn't. Described above. Expected on an alias-mapped endpoint.

Check the operator's status page and changelog before you spend another hour rewriting model names — ours is at status. An endpoint that publishes no model list is an endpoint where these three causes cannot be told apart. You're left guessing spellings for a problem that may not be a spelling at all. That is a property of the endpoint, not of your setup, and it's worth knowing before you build on it.

When reporting it, include the request id — 9Coding error responses carry one in the form (request id: 2026...). That id plus the timestamp, the model name and the full error text lets the exact call be traced. A report that only says it doesn't work cannot be investigated.

404 or 403 — read it as "not available to you"

not_found_error here means not found for this request, which folds together "doesn't exist", "isn't served here" and "isn't yours". Don't read it as proof the model doesn't exist. If you get an explicit 403 with permission_error, that's the same cause B, just stated plainly — see 403 Forbidden.

Related

• 401 Unauthorized — the credential layer, and why rotating the key rarely helps
• 403 Forbidden — authenticated but not permitted
• Context length exceeded — one message, three different limits
• All Claude Code errors — the quick reference table