diff --git a/python/freetoken/server/openai_api.py b/python/freetoken/server/openai_api.py
index 99a4283549..15992930e3 100644
--- a/python/freetoken/server/openai_api.py
+++ b/python/freetoken/server/openai_api.py
@@ -396,6 +396,8 @@ async def handle_completion(
     unsupported = _completion_unsupported_reason(req)
     if unsupported is not None:
         return create_error_response(unsupported)
+    if req.n != 1:
+        return create_error_response("Only n=1 is supported", param="n")
     try:  # surfaces an out-of-range max_tokens as a 400 rather than a 500 from the worker
         default_max_tokens = (
             getattr(state.config, "max_output_tokens", None) or DEFAULT_MAX_OUTPUT_TOKENS
diff --git a/tests/server/test_openai_api.py b/tests/server/test_openai_api.py
index e33018e195..3dcef25f34 100644
--- a/tests/server/test_openai_api.py
+++ b/tests/server/test_openai_api.py
@@ -4,6 +4,7 @@
 import json
 from types import SimpleNamespace
 
+import pytest
 from fastapi import FastAPI
 from fastapi.testclient import TestClient
 from freetoken.message import TokenizeMsg, UserReply
@@ -356,6 +357,24 @@ def test_completion_rejects_token_id_prompts():
     assert "token-id prompt" in body["error"]["message"]
 
 
+@pytest.mark.parametrize("stream", [False, True])
+@pytest.mark.parametrize("n", [0, 2])
+def test_completion_refuses_n_other_than_one(n, stream):
+    state = FakeState([UserReply(uid=42, incremental_output="hello", finished=True)])
+    response = run(
+        handle_completion(
+            CompletionRequest(model="client-model", prompt="say hi", n=n, stream=stream),
+            request=None,
+            state=state,
+            model_sampling={},
+        )
+    )
+
+    assert response.status_code == 400
+    assert json.loads(response.body)["error"]["param"] == "n"
+    assert state.sent is None
+
+
 def test_completion_accepts_text_prompt():
     state = FakeState(
         [UserReply(uid=42, incremental_output="hello", finished=True, prompt_tokens_delta=2, completion_tokens_delta=1)]