diff --git a/src/minisweagent/models/portkey_model.py b/src/minisweagent/models/portkey_model.py
index 6d23aef00..9950a2eb8 100644
--- a/src/minisweagent/models/portkey_model.py
+++ b/src/minisweagent/models/portkey_model.py
@@ -167,6 +167,18 @@ def _calculate_cost(self, response) -> dict[str, float]:
                 f"Completion tokens are None for model {self.config.model_name}. Setting to 0. Full response: {response_for_cost_calc.model_dump()}"
             )
             completion_tokens = 0
+        if total_tokens is None:
+            logger.warning(
+                f"Total tokens are None for model {self.config.model_name}. Setting to sum of prompt and completion tokens. Full response: {response_for_cost_calc.model_dump()}"
+            )
+
+            # Defense: prompt_tokens and completion_tokens are guaranteed to be non-None here
+            # because they were set to 0 above if they were None. But we use get_token() to
+            # be extra safe in case of unexpected API responses.
+            def get_token(val):
+                return val if val is not None else 0
+
+            total_tokens = get_token(prompt_tokens) + get_token(completion_tokens)
         if total_tokens - prompt_tokens - completion_tokens != 0:
             # This is most likely related to how portkey treats cached tokens: It doesn't count them towards the prompt tokens (?)
             logger.warning(
diff --git a/tests/models/test_portkey_model.py b/tests/models/test_portkey_model.py
index f8df5dd78..0b0d13d17 100644
--- a/tests/models/test_portkey_model.py
+++ b/tests/models/test_portkey_model.py
@@ -191,3 +191,52 @@ def test_portkey_model_cost_validation_error():
 
                 assert "Error calculating cost" in str(exc_info.value)
                 assert "MSWEA_COST_TRACKING='ignore_errors'" in str(exc_info.value)
+
+
+def test_portkey_model_total_tokens_none():
+    """Test that None total_tokens is handled gracefully in cost calculation."""
+    mock_portkey_class = MagicMock()
+    mock_client = MagicMock()
+    mock_response = MagicMock()
+    mock_choice = MagicMock()
+    mock_message = MagicMock()
+    mock_usage = MagicMock()
+    mock_tool_call = MagicMock()
+
+    mock_tool_call.id = "call_999"
+    mock_tool_call.function.name = "bash"
+    mock_tool_call.function.arguments = json.dumps({"command": "echo test"})
+    mock_message.tool_calls = [mock_tool_call]
+    mock_message.content = None
+    mock_message.model_dump.return_value = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [{"id": "call_999", "function": {"name": "bash", "arguments": '{"command": "echo test"}'}}],
+    }
+    mock_choice.message = mock_message
+    mock_response.choices = [mock_choice]
+    mock_response.model_dump.return_value = {"test": "response"}
+    mock_response.model_copy.return_value = mock_response
+    mock_response.usage = mock_usage
+    mock_usage.prompt_tokens = 10
+    mock_usage.completion_tokens = 20
+    # Simulate the bug: total_tokens is None
+    mock_usage.total_tokens = None
+
+    mock_client.chat.completions.create.return_value = mock_response
+    mock_portkey_class.return_value = mock_client
+
+    with patch("minisweagent.models.portkey_model.Portkey", mock_portkey_class):
+        with patch.dict(os.environ, {"PORTKEY_API_KEY": "test-key"}):
+            with patch("minisweagent.models.portkey_model.litellm.cost_calculator.completion_cost") as mock_cost:
+                mock_cost.return_value = 0.01
+
+                model = PortkeyModel(model_name="gpt-4o")
+                messages = [{"role": "user", "content": "test"}]
+
+                # This should not raise a TypeError
+                result = model.query(messages)
+
+                # Verify the result is still correct
+                assert result["extra"]["actions"] == [{"command": "echo test", "tool_call_id": "call_999"}]
+                assert result["extra"]["cost"] == 0.01