diff --git a/llm_openai.py b/llm_openai.py
index 4013c31..80ee5fb 100644
--- a/llm_openai.py
+++ b/llm_openai.py
@@ -249,8 +249,6 @@ def _build_messages(self, prompt, conversation):
                                 "arguments": json.dumps(tool_call.arguments),
                             }
                         )
-        if prompt.system and prompt.system != current_system:
-            messages.append({"role": "system", "content": prompt.system})
         if not prompt.attachments:
             messages.append({"role": "user", "content": prompt.prompt or ""})
         else:
@@ -275,6 +273,8 @@ def _build_messages(self, prompt, conversation):
     def _build_kwargs(self, prompt, conversation):
         messages = self._build_messages(prompt, conversation)
         kwargs = {"model": self.model_name, "input": messages}
+        if prompt.system:
+            kwargs["instructions"] = prompt.system
         for option in (
             "max_output_tokens",
             "temperature",
diff --git a/tests/test_openai.py b/tests/test_openai.py
index bb83185..265eb88 100644
--- a/tests/test_openai.py
+++ b/tests/test_openai.py
@@ -76,3 +76,23 @@ def simple_tool(number):
     )
     output = chain_response.text()
     assert output == snapshot
+
+
+def test_instructions_in_kwargs():
+    """Test that system prompt is sent as instructions parameter, not as a system message."""
+    model = llm.get_model("openai/gpt-4o-mini")
+    prompt = llm.Prompt(
+        model=model,
+        prompt="say hi",
+        system="You are a friendly assistant who speaks like a pirate",
+        options=model.Options()
+    )
+    kwargs = model._build_kwargs(prompt, conversation=None)
+
+    # Verify instructions parameter is set
+    assert kwargs["instructions"] == "You are a friendly assistant who speaks like a pirate"
+
+    # Verify no system message in input
+    for item in kwargs["input"]:
+        if isinstance(item, dict):
+            assert item.get("role") != "system"