diff --git a/litellm/llms/oci/chat/transformation.py b/litellm/llms/oci/chat/transformation.py
index aaeb553c176..bca0e5d5754 100644
--- a/litellm/llms/oci/chat/transformation.py
+++ b/litellm/llms/oci/chat/transformation.py
@@ -671,15 +671,25 @@ class OCIChatConfig(BaseConfig):
         Uses litellm's model cost map for a dynamic lookup, falling back to
         DEFAULT_MAX_TOKENS (configurable via the DEFAULT_MAX_TOKENS env var)
         when the model is not found.
+
+        The model cost map stores OCI entries under ``oci/`` prefixed keys
+        (e.g. ``oci/cohere.command-a-03-2025``), but litellm strips the
+        provider prefix before calling ``transform_request``.  We therefore
+        try the bare name first and, if that misses, retry with the
+        ``oci/`` prefix.
         """
         if model is None:
             return DEFAULT_MAX_TOKENS
-        try:
-            max_tokens = get_max_tokens(model)
-            if max_tokens is not None:
-                return max_tokens
-        except Exception:
-            pass
+        # Try the bare model name first, then with the oci/ provider prefix
+        # since the cost map keys are prefixed but transform_request receives
+        # the name without the prefix.
+        for candidate in (model, f"oci/{model}"):
+            try:
+                max_tokens = get_max_tokens(candidate)
+                if max_tokens is not None:
+                    return max_tokens
+            except Exception:
+                continue
         return DEFAULT_MAX_TOKENS
 
     def _get_optional_params(
@@ -907,8 +917,15 @@ class OCIChatConfig(BaseConfig):
         # Build request based on vendor type
         if vendor == OCIVendors.COHERE:
             # For Cohere, we need to use the specific Cohere format
-            # Extract the last user message as the main message
-            user_messages = [msg for msg in messages if msg.get("role") == "user"]
+            # Extract the last user message as the main message.
+            # Filter out user messages with empty content — the Cohere API
+            # requires the main message field to be non-empty.
+            user_messages = [
+                msg
+                for msg in messages
+                if msg.get("role") == "user"
+                and self._extract_text_content(msg.get("content", "")).strip()
+            ]
             if not user_messages:
                 raise Exception("No user message found for Cohere model")
 
diff --git a/tests/test_litellm/llms/oci/chat/test_oci_cohere_max_tokens_and_empty_messages.py b/tests/test_litellm/llms/oci/chat/test_oci_cohere_max_tokens_and_empty_messages.py
index ea4412e82d1..805e983fdaa 100644
--- a/tests/test_litellm/llms/oci/chat/test_oci_cohere_max_tokens_and_empty_messages.py
+++ b/tests/test_litellm/llms/oci/chat/test_oci_cohere_max_tokens_and_empty_messages.py
@@ -74,11 +74,20 @@ class TestOCICohereMaxTokens:
 
         assert params["maxTokens"] == 1024
 
-    def test_get_max_tokens_for_model_known_model(self):
-        """_get_max_tokens_for_model should return the correct value for a known model."""
+    def test_get_max_tokens_for_model_with_oci_prefix(self):
+        """_get_max_tokens_for_model should work with the oci/ prefix."""
         result = OCIChatConfig._get_max_tokens_for_model("oci/cohere.command-a-03-2025")
         assert result == 4000
 
+    def test_get_max_tokens_for_model_without_oci_prefix(self):
+        """_get_max_tokens_for_model should also work without the oci/ prefix.
+
+        litellm strips the provider prefix before calling transform_request,
+        so the method must retry with the oci/ prefix to find the cost map entry.
+        """
+        result = OCIChatConfig._get_max_tokens_for_model("cohere.command-a-03-2025")
+        assert result == 4000
+
     def test_get_max_tokens_for_model_unknown_model(self):
         """_get_max_tokens_for_model should return DEFAULT_MAX_TOKENS for an unknown model."""
         result = OCIChatConfig._get_max_tokens_for_model("oci/unknown-model-xyz")