fix(bedrock): keep cross-region inference-profile prefix for count_tokens

get_bedrock_base_model strips the cross-region inference-profile prefix (global./us./eu./apac./...), so the count-tokens URL used the bare foundation-model ID. For models that are inference-profile-only in a region, Bedrock rejects that with a 400. Resolve the count-tokens/invocation model ID with a new get_bedrock_invocation_model_id helper that preserves the prefix while still stripping routing prefixes, ARNs, and throughput/context-window suffixes.

litellm/llms/bedrock/common_utils.py

@@ -685,6 +685,41 @@ def get_bedrock_base_model(model: str) -> str:
     return model
 
 
+def get_bedrock_invocation_model_id(model: str) -> str:
+    """
+    Resolve the Bedrock model ID for invocation URLs (e.g. ``/model/{id}/converse``
+    or ``/model/{id}/count-tokens``).
+
+    Unlike ``get_bedrock_base_model``, this keeps the cross-region
+    inference-profile prefix (``global.`` / ``us.`` / ``eu.`` / ``apac.`` / ...).
+    Bedrock requires the inference-profile ID for models that are
+    inference-profile-only in a region, so stripping the prefix would produce a
+    bare foundation-model ID that Bedrock rejects with a 400 (see issue #32683).
+    It still strips LiteLLM routing prefixes, resolves ARNs, drops
+    throughput/context-window suffixes, and removes an embedded full-region path
+    prefix (e.g. ``us-east-1/model``).
+    """
+    stripped = model
+    for rp in ["bedrock/converse/", "bedrock/", "converse/"]:
+        if stripped.startswith(rp):
+            stripped = stripped[len(rp) :]
+            break
+    if stripped.startswith("nova-2/"):
+        return "amazon.nova-2-custom"
+    elif stripped.startswith("nova/"):
+        return "amazon.nova-custom"
+
+    model = strip_bedrock_routing_prefix(model)
+    model = extract_model_name_from_bedrock_arn(model)
+    model = strip_bedrock_throughput_suffix(model)
+
+    alt_potential_region = model.split("/", 1)[0]
+    if alt_potential_region in _get_all_bedrock_regions() and len(model.split("/", 1)) > 1:
+        return model.split("/", 1)[1]
+
+    return model
+
+
 def bedrock_converse_supports_parallel_tool_use_config(model: str) -> bool:
     return any(
         (litellm.model_cost.get(candidate) or {}).get("supports_parallel_tool_use_config") is True

litellm/llms/bedrock/count_tokens/bedrock_token_counter.py

@@ -6,7 +6,7 @@ from typing import Any, Dict, List, Optional
 
 from litellm._logging import verbose_logger
 from litellm.llms.base_llm.base_utils import BaseTokenCounter
-from litellm.llms.bedrock.common_utils import BedrockError, get_bedrock_base_model
+from litellm.llms.bedrock.common_utils import BedrockError, get_bedrock_invocation_model_id
 from litellm.llms.bedrock.count_tokens.handler import BedrockCountTokensHandler
 from litellm.types.utils import LlmProviders, TokenCountResponse
 
@@ -68,7 +68,7 @@ class BedrockTokenCounter(BaseTokenCounter):
             request_data["system"] = system
 
         # Get the resolved model (strip prefixes like bedrock/, converse/, etc.)
-        resolved_model = get_bedrock_base_model(model_to_use)
+        resolved_model = get_bedrock_invocation_model_id(model_to_use)
 
         try:
             handler = BedrockCountTokensHandler()

litellm/llms/bedrock/count_tokens/transformation.py

@@ -9,7 +9,7 @@ import re
 from typing import Any, Dict, List, Optional
 
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
-from litellm.llms.bedrock.common_utils import get_bedrock_base_model
+from litellm.llms.bedrock.common_utils import get_bedrock_invocation_model_id
 
 # Placeholder satisfying the Anthropic InvokeModel schema's required
 # max_tokens field; CountTokens only counts input, so it has no effect
@@ -207,12 +207,7 @@ class BedrockCountTokensConfig(BaseAWSLLM):
         Returns:
             Complete endpoint URL for CountTokens API
         """
-        # Use existing LiteLLM function to get the base model ID (removes region prefix)
-        model_id = get_bedrock_base_model(model)
-
-        # Remove bedrock/ prefix if present
-        if model_id.startswith("bedrock/"):
-            model_id = model_id[8:]  # Remove "bedrock/" prefix
+        model_id = get_bedrock_invocation_model_id(model)
         encoded_model_id = self.encode_model_id(model_id=model_id)
 
         base_url, _ = self.get_runtime_endpoint(

tests/litellm_utils_tests/test_bedrock_token_counter.py

@@ -167,6 +167,22 @@ class TestBedrockCountTokensEndpoint:
         )
         assert url == f"{api_base}/model/amazon.nova-lite-v1%3A0/count-tokens"
 
+    def test_cross_region_inference_profile_prefix_preserved(self):
+        """
+        Regression test for #32683: the count-tokens URL must keep the
+        cross-region inference-profile prefix (e.g. global.) so Bedrock does not
+        reject inference-profile-only models with a 400.
+        """
+        handler = self._make_handler()
+        url = handler.get_bedrock_count_tokens_endpoint(
+            model="bedrock/global.anthropic.claude-opus-4-8",
+            aws_region_name="eu-central-1",
+        )
+        assert (
+            url
+            == "https://bedrock-runtime.eu-central-1.amazonaws.com/model/global.anthropic.claude-opus-4-8/count-tokens"
+        )
+
     def test_env_var_overrides_default(self, monkeypatch):
         monkeypatch.setenv(
             "AWS_BEDROCK_RUNTIME_ENDPOINT",

tests/test_litellm/llms/bedrock/test_bedrock_common_utils.py

@@ -285,6 +285,42 @@ def test_context_window_suffix_stripped_for_cost_lookup():
     )
 
 
+def test_get_bedrock_invocation_model_id_preserves_cross_region_prefix():
+    """
+    Regression test for #32683: the count-tokens / invocation model ID must keep
+    the cross-region inference-profile prefix (global./us./eu./apac./...) so
+    Bedrock does not reject inference-profile-only models with a 400.
+    """
+    from litellm.llms.bedrock.common_utils import get_bedrock_invocation_model_id
+
+    assert (
+        get_bedrock_invocation_model_id("bedrock/global.anthropic.claude-opus-4-8")
+        == "global.anthropic.claude-opus-4-8"
+    )
+    assert (
+        get_bedrock_invocation_model_id("eu.anthropic.claude-sonnet-4-6")
+        == "eu.anthropic.claude-sonnet-4-6"
+    )
+    assert (
+        get_bedrock_invocation_model_id("us.meta.llama3-2-11b-instruct-v1:0")
+        == "us.meta.llama3-2-11b-instruct-v1:0"
+    )
+    # Non-prefixed foundation models are unchanged
+    assert (
+        get_bedrock_invocation_model_id("anthropic.claude-opus-4-8")
+        == "anthropic.claude-opus-4-8"
+    )
+    # Routing prefix and throughput/context suffixes are still stripped
+    assert (
+        get_bedrock_invocation_model_id("bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0:51k")
+        == "anthropic.claude-3-5-sonnet-20241022-v2:0"
+    )
+    assert (
+        get_bedrock_invocation_model_id("global.anthropic.claude-opus-4-5-20251101-v1:0[1m]")
+        == "global.anthropic.claude-opus-4-5-20251101-v1:0"
+    )
+
+
 def test_output_config_effort_normalization_uses_model_info_ceiling(monkeypatch):
     import litellm.llms.bedrock.common_utils as mod