fix(bedrock): route overlong GPT version digits to Converse and send native chat completions to the runtime endpoint

A model id with more than 4300 version digits raised ValueError in the route check; the digits are now bounded so such ids fall back to Converse. The native chat completions URL now follows Converse's precedence: aws_bedrock_runtime_endpoint (or AWS_BEDROCK_RUNTIME_ENDPOINT) wins over api_base, so a deployment that sets both keeps sending to the same host

litellm/llms/bedrock/chat/chat_completions/transformation.py

@@ -277,12 +277,12 @@ class AmazonBedrockRuntimeChatCompletionsConfig(OpenAILikeChatConfig):
         aws_region_name: Final = self._aws_signer._get_aws_region_name(  # pyright: ignore[reportPrivateUsage]  # BaseAWSLLM has no public region resolver
             optional_params=self._params_with_region_from_path(optional_params, model), model=model
         )
-        endpoint_url, _ = self._aws_signer.get_runtime_endpoint(
+        _, proxy_endpoint_url = self._aws_signer.get_runtime_endpoint(
             api_base=api_base,
             aws_bedrock_runtime_endpoint=optional_params.get("aws_bedrock_runtime_endpoint"),
             aws_region_name=aws_region_name,
         )
-        base: Final = endpoint_url.rstrip("/")
+        base: Final = proxy_endpoint_url.rstrip("/")
         if base.endswith("/openai/v1/chat/completions"):
             return base
         if base.endswith("/openai/v1"):

litellm/llms/bedrock/common_utils.py

@@ -39,7 +39,7 @@ if TYPE_CHECKING:
 
 _ERROR_REQUEST_URL: Final = "https://docs.litellm.ai/docs"
 _OPENAI_FAMILY_MODEL_RE: Final = re.compile(r"(^|[./])openai\.")
-_OPENAI_GPT_VERSION_RE: Final = re.compile(r"(^|[./])openai\.gpt-(\d+)(?:\.(\d+))?")
+_OPENAI_GPT_VERSION_RE: Final = re.compile(r"(^|[./])openai\.gpt-(\d{1,3})(?!\d)(?:\.(\d{1,3})(?!\d))?")
 _BEDROCK_RUNTIME_CHAT_COMPLETIONS_DEFAULT_SINCE: Final = (5, 6)
 _BEDROCK_RUNTIME_CHAT_COMPLETIONS_ENDPOINT: Final = "/v1/chat/completions"
 BedrockRoute = Literal[

tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py

@@ -149,6 +149,40 @@ def test_complete_url_appends_to_openai_v1_base():
     assert url == "https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1/chat/completions"
 
 
+def test_complete_url_sends_to_the_runtime_endpoint_over_api_base_like_converse(monkeypatch):
+    monkeypatch.delenv("AWS_BEDROCK_RUNTIME_ENDPOINT", raising=False)
+    cfg = AmazonBedrockRuntimeChatCompletionsConfig()
+    url = cfg.get_complete_url(
+        api_base="https://signing-host.example.com",
+        api_key=None,
+        model="us.openai.gpt-5.6-sol",
+        optional_params={"aws_region_name": "us-east-1", "aws_bedrock_runtime_endpoint": "https://egress.example.com/"},
+        litellm_params={},
+    )
+    assert url == "https://egress.example.com/openai/v1/chat/completions"
+
+
+def test_complete_url_sends_to_the_env_runtime_endpoint_over_api_base_like_converse(monkeypatch):
+    monkeypatch.setenv("AWS_BEDROCK_RUNTIME_ENDPOINT", "https://env-egress.example.com")
+    cfg = AmazonBedrockRuntimeChatCompletionsConfig()
+    url = cfg.get_complete_url(
+        api_base="https://signing-host.example.com",
+        api_key=None,
+        model="us.openai.gpt-5.6-sol",
+        optional_params={"aws_region_name": "us-east-1"},
+        litellm_params={},
+    )
+    assert url == "https://env-egress.example.com/openai/v1/chat/completions"
+
+
+@pytest.mark.parametrize("digits", [4, 4301, 30000])
+@pytest.mark.parametrize("template", ["openai.gpt-{run}", "us.openai.gpt-5.{run}", "openai.gpt-{run}.{run}-sol"])
+def test_overlong_gpt_version_digits_route_to_converse_without_raising(local_cost_map, template, digits):
+    model = template.format(run="9" * digits)
+    assert bedrock_runtime_chat_completions_is_default(model) is False
+    assert bedrock_route_for_request(model, {}, None) == "converse"
+
+
 def test_project_id_is_not_sent_as_openai_project_header():
     cfg = AmazonBedrockRuntimeChatCompletionsConfig()
     headers = cfg.validate_environment(