From 6b85e0b4da1d13702929f5e5c14a1784b754f68d Mon Sep 17 00:00:00 2001
From: Jean IBARZ <ibarz.jean@gmail.com>
Date: Fri, 10 Apr 2026 20:46:24 +0200
Subject: [PATCH] fix(together_ai): support reasoning_effort for gpt-oss models

together_ai hosts gpt-oss-120b/20b which support reasoning, but
TogetherAIConfig never advertised reasoning_effort as a supported
OpenAI param. Calling completion(reasoning_effort=...) raised
UnsupportedParamsError even though the underlying model accepts it.

Mirror the deepinfra pattern: read supports_reasoning from the model
info and append reasoning_effort when true. Also set supports_reasoning
on the gpt-oss entries in the model registry so get_model_info reports
the capability.

Fixes #25132
---
 litellm/llms/together_ai/chat.py              |  6 +-
 ...odel_prices_and_context_window_backup.json |  1 +
 model_prices_and_context_window.json          |  1 +
 .../test_together_ai_chat_transformation.py   | 56 +++++++++++++++++++
 4 files changed, 63 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_litellm/llms/together_ai/test_together_ai_chat_transformation.py

diff --git a/litellm/llms/together_ai/chat.py b/litellm/llms/together_ai/chat.py
index e8a784d2779..14206eae107 100644
--- a/litellm/llms/together_ai/chat.py
+++ b/litellm/llms/together_ai/chat.py
@@ -8,8 +8,8 @@ Docs: https://docs.together.ai/reference/completions-1
 
 from typing import Optional
 
-from litellm.utils import get_model_info
 from litellm._logging import verbose_logger
+from litellm.utils import get_model_info
 
 from ..openai.chat.gpt_transformation import OpenAIGPTConfig
 
@@ -22,11 +22,13 @@ class TogetherAIConfig(OpenAIGPTConfig):
         Docs: https://docs.together.ai/docs/json-mode
         """
         supports_function_calling: Optional[bool] = None
+        supports_reasoning: Optional[bool] = None
         try:
             model_info = get_model_info(model, custom_llm_provider="together_ai")
             supports_function_calling = model_info.get(
                 "supports_function_calling", False
             )
+            supports_reasoning = model_info.get("supports_reasoning", False)
         except Exception as e:
             verbose_logger.debug(f"Error getting supported openai params: {e}")
             pass
@@ -40,6 +42,8 @@ class TogetherAIConfig(OpenAIGPTConfig):
             optional_params.remove("tool_choice")
             optional_params.remove("function_call")
             optional_params.remove("response_format")
+        if supports_reasoning is True and "reasoning_effort" not in optional_params:
+            optional_params.append("reasoning_effort")
         return optional_params
 
     def map_openai_params(
diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json
index f7189a60a31..256f35751bf 100644
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@@ -28668,6 +28668,7 @@
         "source": "https://www.together.ai/models/gpt-oss-20b",
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
+        "supports_reasoning": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
     },
diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json
index 3ffc0ec7c58..8d6fe8c8725 100644
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@@ -28653,6 +28653,7 @@
         "source": "https://www.together.ai/models/gpt-oss-20b",
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
+        "supports_reasoning": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
     },
diff --git a/tests/test_litellm/llms/together_ai/test_together_ai_chat_transformation.py b/tests/test_litellm/llms/together_ai/test_together_ai_chat_transformation.py
new file mode 100644
index 00000000000..ff8d4f7a5ff
--- /dev/null
+++ b/tests/test_litellm/llms/together_ai/test_together_ai_chat_transformation.py
@@ -0,0 +1,56 @@
+import os
+import sys
+
+# Add litellm to path
+sys.path.insert(0, os.path.abspath("../../../.."))
+import litellm
+from litellm.llms.together_ai.chat import TogetherAIConfig
+
+
+def _use_local_model_cost_map() -> None:
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+
+def test_together_ai_gpt_oss_supports_reasoning_effort():
+    """
+    gpt-oss models on together_ai support reasoning. Regression test for #25132.
+    """
+    _use_local_model_cost_map()
+
+    supported_params = TogetherAIConfig().get_supported_openai_params(
+        model="together_ai/openai/gpt-oss-120b"
+    )
+    assert "reasoning_effort" in supported_params
+
+    supported_params_20b = TogetherAIConfig().get_supported_openai_params(
+        model="together_ai/openai/gpt-oss-20b"
+    )
+    assert "reasoning_effort" in supported_params_20b
+
+
+def test_together_ai_non_reasoning_model_does_not_expose_reasoning_effort():
+    """
+    Non-reasoning together_ai models should not advertise reasoning_effort.
+    """
+    _use_local_model_cost_map()
+
+    supported_params = TogetherAIConfig().get_supported_openai_params(
+        model="together_ai/meta-llama/Llama-3-70b-chat-hf"
+    )
+    assert "reasoning_effort" not in supported_params
+
+
+def test_together_ai_gpt_oss_reasoning_effort_flows_through_get_optional_params():
+    """
+    End-to-end: the exact call from issue #25132 must not raise
+    UnsupportedParamsError, and reasoning_effort must land in optional_params.
+    """
+    _use_local_model_cost_map()
+
+    optional_params = litellm.get_optional_params(
+        model="openai/gpt-oss-120b",
+        custom_llm_provider="together_ai",
+        reasoning_effort="low",
+    )
+    assert optional_params.get("reasoning_effort") == "low"