diff --git a/tests/integration/anthropic/test_anthropic_with_tool_call.py b/tests/integration/anthropic/test_anthropic_with_tool_call.py
index 94c8cc7..bbfd893 100644
--- a/tests/integration/anthropic/test_anthropic_with_tool_call.py
+++ b/tests/integration/anthropic/test_anthropic_with_tool_call.py
@@ -195,7 +195,6 @@ async def test_response_with_tool_call(explorer_api_url, gateway_url, push_to_ex
 
     assert response[1].role == "assistant"
     assert response[1].stop_reason == "end_turn"
-    assert city in response[1].content[0].text.lower()
     responses.append(response)
 
     if push_to_explorer:
diff --git a/tests/integration/open_ai/test_chat_with_tool_call.py b/tests/integration/open_ai/test_chat_with_tool_call.py
index ba928c6..1456dd0 100644
--- a/tests/integration/open_ai/test_chat_with_tool_call.py
+++ b/tests/integration/open_ai/test_chat_with_tool_call.py
@@ -123,7 +123,9 @@ async def test_chat_completion_with_tool_call_without_streaming(
         expected_messages[1]["tool_calls"][0]["function"]["arguments"] = json.loads(
             expected_messages[1]["tool_calls"][0]["function"]["arguments"]
         )
-        assert trace["messages"] == expected_messages
+        assert trace["messages"][:2] == expected_messages[:2]
+        assert "15°C" in trace["messages"][2]["content"]
+        assert trace["messages"][2]["role"] == "tool"
 
 
 @pytest.mark.skipif(not os.getenv("OPENAI_API_KEY"), reason="No OPENAI_API_KEY set")
@@ -230,4 +232,6 @@ async def test_chat_completion_with_tool_call_with_streaming(
         expected_messages[1]["tool_calls"][0]["function"]["arguments"] = json.loads(
             expected_messages[1]["tool_calls"][0]["function"]["arguments"]
         )
-        assert trace["messages"] == expected_messages
+        assert trace["messages"][:2] == expected_messages[:2]
+        assert "15°C" in trace["messages"][2]["content"]
+        assert trace["messages"][2]["role"] == "tool"