fix(batches): account for Responses API usage

litellm/batches/batch_utils.py

@@ -432,6 +432,10 @@ def _get_batch_job_usage_from_response_body(response_body: dict, custom_llm_prov
             reasoning_content=None,
         )
     _usage_dict = response_body.get("usage", None) or {}
+    from litellm.responses.utils import ResponseAPILoggingUtils
+
+    if ResponseAPILoggingUtils._is_response_api_usage(_usage_dict):
+        return ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(_usage_dict)
     usage: Usage = Usage(**_usage_dict)
     return usage
 

litellm/batches/main.py

@@ -103,7 +103,7 @@ def _resolve_timeout(
 @client
 async def acreate_batch(
     completion_window: Literal["24h"],
-    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions"],
+    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"],
     input_file_id: str,
     custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm"] = "openai",
     metadata: Optional[Dict[str, str]] = None,
@@ -153,7 +153,7 @@ async def acreate_batch(
 @client
 def create_batch(
     completion_window: Literal["24h"],
-    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions"],
+    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"],
     input_file_id: str,
     custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm"] = "openai",
     metadata: Optional[Dict[str, str]] = None,

litellm/types/llms/openai.py

@@ -431,7 +431,7 @@ class CreateBatchRequest(TypedDict, total=False):
     """
 
     completion_window: Literal["24h"]
-    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions"]
+    endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"]
     input_file_id: str
     metadata: Optional[Dict[str, str]]
     output_expires_after: FileExpiresAfter

tests/batches_tests/test_batch_custom_pricing.py

@@ -128,6 +128,29 @@ def test_aggregate_batch_cost_uses_custom_model_info():
     ), f"Expected total cost {expected}, got {cost}"
 
 
+def test_aggregate_batch_cost_normalizes_mixed_responses_and_chat_usage():
+    responses_line = _make_batch_output_line(prompt_tokens=0, completion_tokens=0)
+    responses_line["response"]["body"]["usage"] = {
+        "input_tokens": 20,
+        "output_tokens": 7,
+        "total_tokens": 27,
+        "input_tokens_details": {"cached_tokens": 3},
+    }
+    chat_line = _make_batch_output_line(prompt_tokens=10, completion_tokens=5)
+
+    cost, usage, _ = _aggregate_batch_cost_usage_models(
+        entries=[responses_line, chat_line],
+        custom_llm_provider="openai",
+        model_info=CUSTOM_MODEL_INFO,
+    )
+
+    assert usage.prompt_tokens == 30
+    assert usage.completion_tokens == 12
+    assert usage.total_tokens == 42
+    assert usage.cache_read_input_tokens == 3
+    assert cost == pytest.approx((30 * 0.00125) + (12 * 0.005))
+
+
 @pytest.mark.parametrize("data_residency", ["eu", "us"])
 def test_batch_cost_calculator_applies_data_residency_uplift(
     data_residency, monkeypatch