From f4d84f8f5db0c26e0aa3c50bc2ddb77d07122f5b Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 26 Sep 2026 17:28:06 +0000
Subject: [PATCH] fix(bedrock): drop reasoning_effort none for grok on the
 native chat completions route

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../chat/chat_completions/transformation.py   | 29 ++++++++++++-
 ...bedrock_chat_completions_transformation.py | 42 +++++++++++++++++++
 2 files changed, 70 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/bedrock/chat/chat_completions/transformation.py b/litellm/llms/bedrock/chat/chat_completions/transformation.py
index 4c5e5768119..be2eb8c7713 100644
--- a/litellm/llms/bedrock/chat/chat_completions/transformation.py
+++ b/litellm/llms/bedrock/chat/chat_completions/transformation.py
@@ -60,6 +60,31 @@ def chat_completions_params_refused_for(model: str) -> frozenset[str]:
     )
 
 
+CHAT_COMPLETIONS_REFUSED_REASONING_EFFORTS_BY_FAMILY: Final = MappingProxyType({"xai.": frozenset(("none",))})
+
+
+def chat_completions_reasoning_efforts_refused_for(model: str) -> frozenset[str]:
+    """The ``reasoning_effort`` values AWS's Chat Completions endpoint rejects for this model.
+
+    Grok answers ``"none"`` with a 400 (it takes low, medium, high, and xhigh) where Converse dropped every
+    ``reasoning_effort`` for it, so the native config drops the value and AWS applies its default effort as before.
+    """
+    model_id: Final = split_bedrock_region_path(model)[1]
+    return frozenset().union(
+        *(
+            refused
+            for family, refused in CHAT_COMPLETIONS_REFUSED_REASONING_EFFORTS_BY_FAMILY.items()
+            if family in model_id
+        )
+    )
+
+
+def without_refused_reasoning_effort(model: str, params: Mapping[str, object]) -> Mapping[str, object]:
+    if params.get("reasoning_effort") not in chat_completions_reasoning_efforts_refused_for(model):
+        return params
+    return MappingProxyType({key: value for key, value in params.items() if key != "reasoning_effort"})
+
+
 def _held_close_tag_prefix(text: str) -> int:
     return next(
         (
@@ -277,7 +302,9 @@ class AmazonBedrockRuntimeChatCompletionsConfig(OpenAILikeChatConfig):
             drop_params=drop_params,
             replace_max_completion_tokens_with_max_tokens=replace_max_completion_tokens_with_max_tokens,
         )
-        return dict(with_max_completion_tokens(mapped))  # mutable-ok: get_optional_params keeps filling this dict
+        return dict(  # mutable-ok: get_optional_params keeps filling this dict
+            without_refused_reasoning_effort(model, with_max_completion_tokens(mapped))
+        )
 
     def _inference_params(
         self, optional_params: Mapping[str, object]
diff --git a/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py b/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
index 3c347e4bed2..061df201c15 100644
--- a/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
+++ b/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
@@ -11,6 +11,7 @@ from litellm.llms.bedrock.chat.chat_completions.transformation import (
     AmazonBedrockRuntimeChatCompletionsConfig,
     BedrockRuntimeChatCompletionsStreamingHandler,
     ReasoningTagSplitter,
+    chat_completions_reasoning_efforts_refused_for,
     split_reasoning_tag,
     with_max_completion_tokens,
 )
@@ -333,6 +334,47 @@ def test_with_max_completion_tokens_leaves_other_params_alone():
     assert with_max_completion_tokens({"temperature": 0.5}) == {"temperature": 0.5}
 
 
+@pytest.mark.parametrize(
+    "model",
+    ["us.xai.grok-4.6", "bedrock/us-gov-west-1/us.xai.grok-4.6"],
+)
+def test_map_openai_params_drops_reasoning_effort_none_for_grok(model):
+    cfg = AmazonBedrockRuntimeChatCompletionsConfig()
+    mapped = cfg.map_openai_params(
+        non_default_params={"reasoning_effort": "none", "max_tokens": 64},
+        optional_params={},
+        model=model,
+        drop_params=False,
+    )
+    assert "reasoning_effort" not in mapped
+
+
+def test_map_openai_params_keeps_reasoning_effort_low_for_grok():
+    cfg = AmazonBedrockRuntimeChatCompletionsConfig()
+    mapped = cfg.map_openai_params(
+        non_default_params={"reasoning_effort": "low", "max_tokens": 64},
+        optional_params={},
+        model="us.xai.grok-4.6",
+        drop_params=False,
+    )
+    assert mapped["reasoning_effort"] == "low"
+
+
+def test_map_openai_params_keeps_reasoning_effort_none_for_gpt56():
+    cfg = AmazonBedrockRuntimeChatCompletionsConfig()
+    mapped = cfg.map_openai_params(
+        non_default_params={"reasoning_effort": "none", "max_tokens": 64},
+        optional_params={},
+        model="global.openai.gpt-5.6-sol",
+        drop_params=False,
+    )
+    assert mapped["reasoning_effort"] == "none"
+
+
+def test_reasoning_efforts_refused_for_is_empty_outside_xai():
+    assert chat_completions_reasoning_efforts_refused_for("openai.gpt-oss-20b-1:0") == frozenset()
+
+
 def test_supported_params_include_reasoning_effort_for_gpt56(local_cost_map):
     cfg = AmazonBedrockRuntimeChatCompletionsConfig()
     assert "reasoning_effort" in cfg.get_supported_openai_params("global.openai.gpt-5.6-sol")