diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 6b559b9c45..037606ced8 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -2797,6 +2797,13 @@ def _normalise_responses_input(payload: ResponsesRequest) -> list[ChatMessage]:
             output = item.output
             if not isinstance(output, str):
                 output = json.dumps(output)
+            if not output:
+                # An empty/falsy result (e.g. an image-only tool output whose
+                # payload lives outside `output`) would otherwise build a
+                # role="tool" ChatMessage with content="", which the strict
+                # validator rejects with a 500. Emit a placeholder so the turn
+                # normalises gracefully.
+                output = "(no output)"
             messages.append(
                 ChatMessage(
                     role = "tool",
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index 2f1161c329..be99454483 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -374,6 +374,44 @@ class TestNormaliseResponsesInputWithTools:
         # Content is serialised so llama-server sees a string.
         assert json.loads(msgs[0].content) == [{"type": "output_text", "text": "ok"}]
 
+    def test_empty_tool_output_replaced_with_placeholder(self):
+        """Image-only tool results (Anthropic format) send empty output.
+        Before the fix this crashed the validator with
+        "role=\"tool\" messages require non-empty \"content\"".
+        Now it emits "(no output)" so the turn normalises cleanly."""
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": "",
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].role == "tool"
+        assert msgs[0].tool_call_id == "call_1"
+        assert msgs[0].content == "(no output)"
+        # Round-trip through ChatMessage validator must not raise.
+        ChatMessage(**msgs[0].model_dump(exclude_none = True))
+
+    def test_empty_list_output_serialised_to_json_array(self):
+        """Empty list output is not falsy after json.dumps (becomes "[]"),
+        so it should not trigger the placeholder."""
+        payload = ResponsesRequest(
+            input = [
+                {
+                    "type": "function_call_output",
+                    "call_id": "call_1",
+                    "output": [],
+                }
+            ],
+        )
+        msgs = _normalise_responses_input(payload)
+        assert msgs[0].role == "tool"
+        assert msgs[0].content == "[]"
+        ChatMessage(**msgs[0].model_dump(exclude_none = True))
+
 
 # =====================================================================
 # Response mapping — tool_calls → function_call output items