diff --git a/litellm/llms/oci/chat/cohere.py b/litellm/llms/oci/chat/cohere.py
index ce3560cd2bb..39caebba03a 100644
--- a/litellm/llms/oci/chat/cohere.py
+++ b/litellm/llms/oci/chat/cohere.py
@@ -294,6 +294,12 @@ def handle_cohere_stream_chunk(dict_chunk: dict) -> ModelResponseStream:
         finish_reason = "length"
     elif finish_reason == "TOOL_CALL":
         finish_reason = "tool_calls"
+    elif finish_reason is not None:
+        # OCI Cohere can emit error/cancel finish reasons (e.g. ``ERROR``,
+        # ``ERROR_TOXIC``, ``ERROR_LIMIT``, ``USER_CANCEL``) that aren't part
+        # of OpenAI's standard set. Normalize them to ``"stop"`` so downstream
+        # consumers switching on ``finish_reason`` keep working.
+        finish_reason = "stop"
 
     return ModelResponseStream(
         choices=[
diff --git a/litellm/llms/oci/chat/transformation.py b/litellm/llms/oci/chat/transformation.py
index f8ecc479cae..1bf960e5efc 100644
--- a/litellm/llms/oci/chat/transformation.py
+++ b/litellm/llms/oci/chat/transformation.py
@@ -209,6 +209,12 @@ class OCIChatConfig(BaseConfig):
         # the param dropped under drop_params, rather than silently passing
         # through and tripping Pydantic validation on CohereChatRequest.
         self.openai_to_oci_cohere_param_map["reasoning_effort"] = False
+        # CohereChatRequest has no logProbs/logitBias fields, so passing these
+        # through would be silently dropped by Pydantic. Mark them unsupported
+        # so get_supported_openai_params doesn't advertise them and callers
+        # get a clear error (or drop_params behaviour) instead.
+        self.openai_to_oci_cohere_param_map["logprobs"] = False
+        self.openai_to_oci_cohere_param_map["logit_bias"] = False
 
     def get_supported_openai_params(self, model: str) -> List[str]:
         param_map = (