diff --git a/src/ra/clients/openai.py b/src/ra/clients/openai.py
index 2f700f42..f78fefcf 100644
--- a/src/ra/clients/openai.py
+++ b/src/ra/clients/openai.py
@@ -6,6 +6,15 @@
 from dotenv import load_dotenv
 from openai.types.chat import ChatCompletion
 
+# Use Langfuse-wrapped OpenAI client when available for automatic tracing
+_LangfuseOpenAI = None
+_LangfuseAsyncOpenAI = None
+if os.environ.get("LANGFUSE_SECRET_KEY"):
+    try:
+        from langfuse.openai import OpenAI as _LangfuseOpenAI, AsyncOpenAI as _LangfuseAsyncOpenAI
+    except ImportError:
+        pass
+
 from ra.clients.base_lm import BaseLM
 from ra.core.types import ModelUsageSummary, UsageSummary
 
@@ -51,7 +60,8 @@ def __init__(
         extra_headers = (
             _OPENROUTER_HEADERS if base_url == "https://openrouter.ai/api/v1" else None
         )
-        self.client = openai.OpenAI(
+        _ClientCls = _LangfuseOpenAI or openai.OpenAI
+        self.client = _ClientCls(
             api_key=api_key,
             base_url=base_url,
             default_headers=extra_headers,
@@ -121,7 +131,8 @@ async def acompletion(
         if self.client.base_url == DEFAULT_PRIME_INTELLECT_BASE_URL:
             extra_body["usage"] = {"include": True}
 
-        async with openai.AsyncOpenAI(
+        _AsyncCls = _LangfuseAsyncOpenAI or openai.AsyncOpenAI
+        async with _AsyncCls(
             **self._async_client_kwargs,
         ) as client:
             response = await client.chat.completions.create(