fix(bedrock): strip cross-region inference profile prefixes for context window

lib/crewai/src/crewai/llms/providers/bedrock/completion.py

@@ -2090,6 +2090,20 @@ def supports_stop_words(self) -> bool:
         """Check if the model supports stop words."""
         return True
 
+    @staticmethod
+    def _strip_inference_profile_prefix(model_id: str) -> str:
+        """Strip geographic cross-region inference profile prefixes.
+
+        AWS Bedrock inference profile ids look like
+        ``us.anthropic.claude-sonnet-4-...`` / ``eu.`` / ``apac.`` / ``us-gov.``.
+        Context-window tables are keyed on the bare foundation-model id.
+        """
+        model_id = model_id.removeprefix("bedrock/")
+        for geo in ("us-gov.", "apac.", "us.", "eu.", "jp.", "au.", "ca."):
+            if model_id.startswith(geo):
+                return model_id[len(geo) :]
+        return model_id
+
     def get_context_window_size(self) -> int:
         """Get the context window size for the model."""
         from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO
@@ -2114,8 +2128,12 @@ def get_context_window_size(self) -> int:
             "deepseek.r1": 32768,
         }
 
-        for model_prefix, size in context_windows.items():
-            if self.model.startswith(model_prefix):
+        model_id = self._strip_inference_profile_prefix(self.model)
+        # Longest prefix first so claude-3-5-sonnet wins over claude-3
+        for model_prefix, size in sorted(
+            context_windows.items(), key=lambda item: len(item[0]), reverse=True
+        ):
+            if model_id.startswith(model_prefix):
                 return int(size * CONTEXT_WINDOW_USAGE_RATIO)
 
         return int(8192 * CONTEXT_WINDOW_USAGE_RATIO)

lib/crewai/tests/llms/bedrock/test_bedrock.py

@@ -530,6 +530,21 @@ def test_bedrock_context_window_size():
     assert context_size_titan > 5000
 
 
+def test_bedrock_cross_region_inference_profile_context_window():
+    """Cross-region inference profile ids (us./eu./apac./us-gov.) must map to the bare model window."""
+    from crewai.llm import CONTEXT_WINDOW_USAGE_RATIO
+
+    expected = int(200000 * CONTEXT_WINDOW_USAGE_RATIO)
+    for model in (
+        "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0",
+        "bedrock/eu.anthropic.claude-sonnet-4-20250514-v1:0",
+        "bedrock/apac.anthropic.claude-3-5-sonnet-20241022-v2:0",
+        "us.anthropic.claude-sonnet-4-20250514-v1:0",
+    ):
+        llm = LLM(model=model)
+        assert llm.get_context_window_size() == expected, model
+
+
 def test_bedrock_message_formatting():
     """
     Test that messages are properly formatted for Bedrock Converse API