diff --git a/litellm/llms/custom_httpx/async_client_cleanup.py b/litellm/llms/custom_httpx/async_client_cleanup.py
index 9c1f6af7e9c..f7f5c90d955 100644
--- a/litellm/llms/custom_httpx/async_client_cleanup.py
+++ b/litellm/llms/custom_httpx/async_client_cleanup.py
@@ -3,6 +3,7 @@ Utility functions for cleaning up async HTTP clients to prevent resource leaks.
 """
 
 import asyncio
+import inspect
 
 
 async def close_litellm_async_clients():
@@ -46,13 +47,22 @@ async def close_litellm_async_clients():
                     # Silently ignore errors during cleanup
                     pass
 
-        # Handle any other objects with aclose method
-        elif hasattr(handler, "aclose"):
-            try:
-                await handler.aclose()
-            except Exception:
-                # Silently ignore errors during cleanup
-                pass
+        # Handle any other cached async client. Most expose aclose(), but some
+        # SDK clients -- notably the openai SDK's AsyncOpenAI / AsyncAzureOpenAI,
+        # which are cached here (set_cached_openai_client) and built on litellm's
+        # aiohttp transport -- name the async teardown close() instead. Accept
+        # either, guarding on iscoroutinefunction so sync close() methods
+        # (sync OpenAI / HTTPHandler) are left untouched.
+        else:
+            closer = getattr(handler, "aclose", None) or getattr(
+                handler, "close", None
+            )
+            if closer is not None and inspect.iscoroutinefunction(closer):
+                try:
+                    await closer()
+                except Exception:
+                    # Silently ignore errors during cleanup
+                    pass
 
     # Close the global base_llm_aiohttp_handler instance (issue #12443)
     # This is used by Gemini and other providers that use aiohttp
diff --git a/tests/test_litellm/llms/custom_httpx/test_openai_session_leak.py b/tests/test_litellm/llms/custom_httpx/test_openai_session_leak.py
new file mode 100644
index 00000000000..fde184626ac
--- /dev/null
+++ b/tests/test_litellm/llms/custom_httpx/test_openai_session_leak.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+"""
+Test for the AsyncOpenAI cleanup gap: close_litellm_async_clients() must close
+cached clients whose async teardown is named close() (the openai SDK's
+AsyncOpenAI / AsyncAzureOpenAI), not only those exposing aclose().
+
+Proposed path: tests/test_litellm/llms/custom_httpx/test_openai_session_leak.py
+"""
+
+import pytest
+
+import litellm
+from litellm.llms.custom_httpx.async_client_cleanup import close_litellm_async_clients
+
+
+class _AsyncCloseOnlyClient:
+    """Mirrors openai.AsyncOpenAI: async close(), no aclose(), no .client."""
+
+    def __init__(self):
+        self.closed = False
+
+    async def close(self):
+        self.closed = True
+
+
+class _SyncClient:
+    """A sync client (close() is not a coroutine) -- must be left untouched."""
+
+    def __init__(self):
+        self.closed = False
+
+    def close(self):
+        self.closed = True
+
+
+@pytest.mark.asyncio
+async def test_cleanup_closes_async_close_only_clients():
+    async_client = _AsyncCloseOnlyClient()
+    sync_client = _SyncClient()
+    litellm.in_memory_llm_clients_cache.set_cache(
+        key="test_async_close_only", value=async_client, ttl=60
+    )
+    litellm.in_memory_llm_clients_cache.set_cache(
+        key="test_sync_close", value=sync_client, ttl=60
+    )
+
+    await close_litellm_async_clients()
+
+    assert async_client.closed, "async close()-only client should be closed"
+    assert not sync_client.closed, "sync client should be left untouched"