diff --git a/litellm/integrations/websearch_interception/handler.py b/litellm/integrations/websearch_interception/handler.py
index 1db94e82066..8ffdd5f4b4f 100644
--- a/litellm/integrations/websearch_interception/handler.py
+++ b/litellm/integrations/websearch_interception/handler.py
@@ -1284,16 +1284,25 @@ class WebSearchInterceptionLogger(CustomLogger):
     def _resolve_max_tokens(
         optional_params: dict,
         kwargs: dict,
+        logging_obj: "LiteLLMLoggingObj | None" = None,
     ) -> int:
         """Extract max_tokens and validate against thinking.budget_tokens.
 
         Anthropic API requires ``max_tokens > thinking.budget_tokens``.
         If the constraint is violated, auto-adjust to ``budget_tokens + 1024``.
         """
-        max_tokens: int = optional_params.get(
-            "max_tokens",
-            kwargs.get("max_tokens", 1024),
-        )
+        # Look in optional_params, kwargs, or the original model_call_details / litellm_params
+        max_tokens_val = optional_params.get("max_tokens")
+        if max_tokens_val is None:
+            max_tokens_val = kwargs.get("max_tokens")
+        if max_tokens_val is None and logging_obj is not None:
+            model_details = getattr(logging_obj, "model_call_details", {}) or {}
+            max_tokens_val = model_details.get("max_tokens")
+            if max_tokens_val is None:
+                litellm_params = model_details.get("litellm_params", {}) or {}
+                max_tokens_val = litellm_params.get("max_tokens")
+
+        max_tokens: int = max_tokens_val if max_tokens_val is not None else 1024
         thinking_param: Final = optional_params.get("thinking")
         if thinking_param and isinstance(thinking_param, dict):
             budget_tokens: Final = thinking_param.get("budget_tokens")
@@ -1437,7 +1446,9 @@ class WebSearchInterceptionLogger(CustomLogger):
 
         full_model_name = model  # safe default before try block
 
-        max_tokens: Final = self._resolve_max_tokens(anthropic_messages_optional_request_params, kwargs)
+        max_tokens: Final = self._resolve_max_tokens(
+            anthropic_messages_optional_request_params, kwargs, logging_obj=logging_obj
+        )
 
         verbose_logger.debug("WebSearchInterception: Using max_tokens=%s for follow-up request", max_tokens)
 
diff --git a/tests/unit/integrations/websearch_interception/test_websearch_thinking_constraint.py b/tests/unit/integrations/websearch_interception/test_websearch_thinking_constraint.py
index b2d5225070c..47082324a0f 100644
--- a/tests/unit/integrations/websearch_interception/test_websearch_thinking_constraint.py
+++ b/tests/unit/integrations/websearch_interception/test_websearch_thinking_constraint.py
@@ -491,3 +491,38 @@ class TestFollowUpErrorScenarios:
         # But ALL proxy metadata must be preserved
         assert captured_kwargs.get("metadata") == proxy_metadata
         assert captured_kwargs.get("litellm_call_id") == "call-abc-123"
+
+    @pytest.mark.asyncio
+    async def test_max_tokens_resolved_from_logging_obj_when_popped_from_optional_params(self):
+        """When max_tokens was popped from optional_params (Messages transformation), it is resolved from logging_obj."""
+        logger = WebSearchInterceptionLogger(enabled_providers=["bedrock"])
+        captured_kwargs: Dict[str, Any] = {}
+
+        async def _fake_acreate(**kw):
+            captured_kwargs.update(kw)
+            return MagicMock()
+
+        logging_obj = _make_logging_obj()
+        logging_obj.model_call_details["max_tokens"] = 4000
+
+        with (
+            patch(
+                "litellm.integrations.websearch_interception.handler.anthropic_messages.acreate",
+                side_effect=_fake_acreate,
+            ),
+            patch.object(
+                logger, "_execute_search", return_value=("search result", None)
+            ),
+        ):
+            await logger._execute_agentic_loop(
+                model="us.anthropic.claude-opus-4-6-v1",
+                messages=[{"role": "user", "content": "hi"}],
+                tool_calls=_make_tool_calls(),
+                thinking_blocks=[],
+                anthropic_messages_optional_request_params={},  # max_tokens already popped by transformation
+                logging_obj=logging_obj,
+                stream=False,
+                kwargs={},
+            )
+
+        assert captured_kwargs.get("max_tokens") == 4000