diff --git a/.sampo/changesets/ai-tool-call-spans.md b/.sampo/changesets/ai-tool-call-spans.md
new file mode 100644
index 00000000..50e65baa
--- /dev/null
+++ b/.sampo/changesets/ai-tool-call-spans.md
@@ -0,0 +1,5 @@
+---
+pypi/posthog: minor
+---
+
+feat(ai): emit an `$ai_span` for each tool call an LLM requests in the OpenAI, Anthropic, and Gemini wrappers. When a generation's response contains tool calls, the wrapper now captures a child span (`$ai_span_type: "tool"`) per call — carrying the tool name, arguments, and tool-call id — nested under the generation via `$ai_parent_id`, so requested tools show up in the trace tree. Generation events also now carry an `$ai_span_id`. Works for sync, async, and streaming calls; span capture is best-effort and never breaks the underlying LLM call.
diff --git a/posthog/ai/utils.py b/posthog/ai/utils.py
index ec369031..6c1fb2cd 100644
--- a/posthog/ai/utils.py
+++ b/posthog/ai/utils.py
@@ -379,6 +379,88 @@ def merge_system_prompt(
     return []
 
 
+def _extract_response_tool_calls(formatted_output: Any) -> List[Dict[str, Any]]:
+    """Pull the tool calls the model requested out of an already-formatted response.
+
+    Every provider converter normalizes a tool call into a content item shaped
+    like ``{"type": "function", "id": ..., "function": {"name": ...,
+    "arguments": ...}}`` (see the ``format_*_output`` helpers in each
+    converter), so reading that shape keeps this provider agnostic.
+    """
+    tool_calls: List[Dict[str, Any]] = []
+    if not isinstance(formatted_output, list):
+        return tool_calls
+    for message in formatted_output:
+        if not isinstance(message, dict):
+            continue
+        content = message.get("content")
+        if not isinstance(content, list):
+            continue
+        for item in content:
+            if isinstance(item, dict) and item.get("type") == "function":
+                fn = item.get("function") or {}
+                tool_calls.append(
+                    {
+                        "id": item.get("id"),
+                        "name": fn.get("name"),
+                        "arguments": fn.get("arguments"),
+                    }
+                )
+    return tool_calls
+
+
+def _capture_tool_call_spans(
+    ph_client: PostHogClient,
+    *,
+    formatted_output: Any,
+    trace_id: str,
+    parent_id: str,
+    distinct_id: Optional[str],
+    groups: Optional[Dict[str, Any]],
+    privacy_mode: bool,
+    include_person_profile: bool,
+) -> None:
+    """Emit one ``$ai_span`` per tool call the LLM requested in its response.
+
+    Each span records *which* tool the model chose (name + arguments), nested
+    under the generation via ``$ai_parent_id`` so it shows up in the trace tree.
+    The wrapper never runs the tool, so these spans carry no duration or result
+    — instrument the tool execution itself if you need that. Best-effort: a
+    failure here must never break the underlying LLM call.
+    """
+    try:
+        tool_calls = _extract_response_tool_calls(formatted_output)
+        if not tool_calls:
+            return
+        for tool_call in tool_calls:
+            properties: Dict[str, Any] = {
+                "$ai_trace_id": trace_id,
+                "$ai_span_id": str(uuid.uuid4()),
+                "$ai_parent_id": parent_id,
+                "$ai_span_name": tool_call.get("name") or "tool_call",
+                "$ai_span_type": "tool",
+                "$ai_input_state": with_privacy_mode(
+                    ph_client, privacy_mode, tool_call.get("arguments")
+                ),
+                "$ai_latency": 0,
+            }
+            tool_call_id = tool_call.get("id")
+            if tool_call_id is not None:
+                properties["$ai_tool_call_id"] = tool_call_id
+            if not include_person_profile:
+                properties["$process_person_profile"] = False
+            _capture_ai_event(
+                ph_client,
+                "$ai_span",
+                distinct_id=distinct_id,
+                properties=properties,
+                groups=groups,
+            )
+    except Exception:
+        # Span capture is best-effort telemetry; never surface to the caller.
+        pass
+
+
 def call_llm_and_track_usage(
     posthog_distinct_id: Optional[str],
     ph_client: PostHogClient,
@@ -453,19 +535,20 @@ def call_llm_and_track_usage(
                 "$ai_input",
                 with_privacy_mode(ph_client, posthog_privacy_mode, sanitized_messages),
             )
+            formatted_output = finalize_ai_content(
+                format_response(response, provider), ph_client
+            )
             tag(
                 "$ai_output_choices",
-                with_privacy_mode(
-                    ph_client,
-                    posthog_privacy_mode,
-                    finalize_ai_content(format_response(response, provider), ph_client),
-                ),
+                with_privacy_mode(ph_client, posthog_privacy_mode, formatted_output),
             )
             tag("$ai_http_status", http_status)
             tag("$ai_input_tokens", usage.get("input_tokens", 0))
             tag("$ai_output_tokens", usage.get("output_tokens", 0))
             tag("$ai_latency", latency)
             tag("$ai_trace_id", posthog_trace_id)
+            generation_span_id = str(uuid.uuid4())
+            tag("$ai_span_id", generation_span_id)
             tag("$ai_base_url", str(base_url))
             warn_if_posthog_ai_gateway(base_url)
 
@@ -529,6 +612,16 @@ def call_llm_and_track_usage(
                     properties=merged_properties,
                     groups=posthog_groups,
                 )
+                _capture_tool_call_spans(
+                    ph_client,
+                    formatted_output=formatted_output,
+                    trace_id=posthog_trace_id,
+                    parent_id=generation_span_id,
+                    distinct_id=contexts.get_context_distinct_id(),
+                    groups=posthog_groups,
+                    privacy_mode=posthog_privacy_mode,
+                    include_person_profile=has_person_distinct_id,
+                )
 
         if error:
             raise error
@@ -606,19 +699,20 @@ async def call_llm_and_track_usage_async(
                 "$ai_input",
                 with_privacy_mode(ph_client, posthog_privacy_mode, sanitized_messages),
             )
+            formatted_output = finalize_ai_content(
+                format_response(response, provider), ph_client
+            )
             tag(
                 "$ai_output_choices",
-                with_privacy_mode(
-                    ph_client,
-                    posthog_privacy_mode,
-                    finalize_ai_content(format_response(response, provider), ph_client),
-                ),
+                with_privacy_mode(ph_client, posthog_privacy_mode, formatted_output),
             )
             tag("$ai_http_status", http_status)
             tag("$ai_input_tokens", usage.get("input_tokens", 0))
             tag("$ai_output_tokens", usage.get("output_tokens", 0))
             tag("$ai_latency", latency)
             tag("$ai_trace_id", posthog_trace_id)
+            generation_span_id = str(uuid.uuid4())
+            tag("$ai_span_id", generation_span_id)
             tag("$ai_base_url", str(base_url))
             warn_if_posthog_ai_gateway(base_url)
 
@@ -682,6 +776,16 @@ async def call_llm_and_track_usage_async(
                     properties=merged_properties,
                     groups=posthog_groups,
                 )
+                _capture_tool_call_spans(
+                    ph_client,
+                    formatted_output=formatted_output,
+                    trace_id=posthog_trace_id,
+                    parent_id=generation_span_id,
+                    distinct_id=contexts.get_context_distinct_id(),
+                    groups=posthog_groups,
+                    privacy_mode=posthog_privacy_mode,
+                    include_person_profile=has_person_distinct_id,
+                )
 
         if error:
             raise error
@@ -728,6 +832,7 @@ def capture_streaming_event(
         event_data: Standardized streaming event data containing all necessary information
     """
     trace_id = event_data.get("trace_id") or str(uuid.uuid4())
+    generation_span_id = str(uuid.uuid4())
 
     # Build base event properties
     event_properties = {
@@ -749,6 +854,7 @@ def capture_streaming_event(
         "$ai_output_tokens": event_data["usage_stats"].get("output_tokens", 0),
         "$ai_latency": event_data["latency"],
         "$ai_trace_id": trace_id,
+        "$ai_span_id": generation_span_id,
         "$ai_base_url": str(event_data["base_url"]),
         **(event_data.get("properties") or {}),
     }
@@ -837,3 +943,13 @@ def capture_streaming_event(
             properties=event_properties,
             groups=event_data.get("groups"),
         )
+        _capture_tool_call_spans(
+            ph_client,
+            formatted_output=event_data["formatted_output"],
+            trace_id=trace_id,
+            parent_id=generation_span_id,
+            distinct_id=event_data.get("distinct_id") or trace_id,
+            groups=event_data.get("groups"),
+            privacy_mode=event_data["privacy_mode"],
+            include_person_profile=event_data.get("distinct_id") is not None,
+        )
diff --git a/posthog/test/ai/anthropic/test_anthropic.py b/posthog/test/ai/anthropic/test_anthropic.py
index 62cdc24d..d5be7d0e 100644
--- a/posthog/test/ai/anthropic/test_anthropic.py
+++ b/posthog/test/ai/anthropic/test_anthropic.py
@@ -917,9 +917,14 @@ def test_tool_calls_in_output_choices(
         )
 
         assert response == mock_anthropic_response_with_tool_calls
-        assert mock_client.capture.call_count == 1
+        # $ai_generation plus one $ai_span for the tool call the model requested
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        call_args = next(
+            c[1]
+            for c in mock_client.capture.call_args_list
+            if c[1]["event"] == "$ai_generation"
+        )
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -980,9 +985,14 @@ def test_tool_calls_only_no_content(
         )
 
         assert response == mock_anthropic_response_tool_calls_only
-        assert mock_client.capture.call_count == 1
+        # $ai_generation plus one $ai_span for the requested tool call
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        call_args = next(
+            c[1]
+            for c in mock_client.capture.call_args_list
+            if c[1]["event"] == "$ai_generation"
+        )
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -1049,9 +1059,14 @@ async def run_test():
         response = asyncio.run(run_test())
 
         assert response == mock_anthropic_response_with_tool_calls
-        assert mock_client.capture.call_count == 1
+        # $ai_generation plus one $ai_span for the requested tool call
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        call_args = next(
+            c[1]
+            for c in mock_client.capture.call_args_list
+            if c[1]["event"] == "$ai_generation"
+        )
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -1117,9 +1132,14 @@ def test_streaming_with_tool_calls(mock_client, mock_anthropic_stream_with_tools
         list(response)
 
         # Capture happens synchronously when generator is exhausted
-        assert mock_client.capture.call_count == 1
+        # ($ai_generation plus one $ai_span for the streamed tool call)
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        call_args = next(
+            c[1]
+            for c in mock_client.capture.call_args_list
+            if c[1]["event"] == "$ai_generation"
+        )
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -1244,9 +1264,14 @@ async def run_test():
         asyncio.run(run_test())
 
         # Capture completes before asyncio.run() returns
-        assert mock_client.capture.call_count == 1
+        # ($ai_generation plus one $ai_span for the streamed tool call)
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        call_args = next(
+            c[1]
+            for c in mock_client.capture.call_args_list
+            if c[1]["event"] == "$ai_generation"
+        )
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
diff --git a/posthog/test/ai/gemini/test_gemini.py b/posthog/test/ai/gemini/test_gemini.py
index c1b82fe0..dcaf93b0 100644
--- a/posthog/test/ai/gemini/test_gemini.py
+++ b/posthog/test/ai/gemini/test_gemini.py
@@ -718,9 +718,14 @@ def test_function_calls_in_output_choices(
     )
 
     assert response == mock_gemini_response_with_function_calls
-    assert mock_client.capture.call_count == 1
+    # $ai_generation plus one $ai_span for the function call the model requested
+    assert mock_client.capture.call_count == 2
 
-    call_args = mock_client.capture.call_args[1]
+    call_args = next(
+        c[1]
+        for c in mock_client.capture.call_args_list
+        if c[1]["event"] == "$ai_generation"
+    )
     props = call_args["properties"]
 
     assert call_args["distinct_id"] == "test-id"
@@ -767,9 +772,14 @@ def test_function_calls_only_no_content(
     )
 
     assert response == mock_gemini_response_function_calls_only
-    assert mock_client.capture.call_count == 1
+    # $ai_generation plus one $ai_span for the requested function call
+    assert mock_client.capture.call_count == 2
 
-    call_args = mock_client.capture.call_args[1]
+    call_args = next(
+        c[1]
+        for c in mock_client.capture.call_args_list
+        if c[1]["event"] == "$ai_generation"
+    )
     props = call_args["properties"]
 
     assert call_args["distinct_id"] == "test-id"
@@ -1553,8 +1563,13 @@ def mock_streaming_response():
 
     list(response)
 
-    assert mock_client.capture.call_count == 1
-    props = mock_client.capture.call_args[1]["properties"]
+    # $ai_generation plus one $ai_span for the streamed function call
+    assert mock_client.capture.call_count == 2
+    props = next(
+        c[1]
+        for c in mock_client.capture.call_args_list
+        if c[1]["event"] == "$ai_generation"
+    )["properties"]
     assert props["$ai_stop_reason"] == "STOP"
 
 
diff --git a/posthog/test/ai/gemini/test_gemini_async.py b/posthog/test/ai/gemini/test_gemini_async.py
index 53f85b26..a10ca291 100644
--- a/posthog/test/ai/gemini/test_gemini_async.py
+++ b/posthog/test/ai/gemini/test_gemini_async.py
@@ -586,9 +586,14 @@ async def test_async_function_calls_in_output_choices(
     )
 
     assert response == mock_gemini_response_with_function_calls
-    assert mock_client.capture.call_count == 1
+    # $ai_generation plus one $ai_span for the function call the model requested
+    assert mock_client.capture.call_count == 2
 
-    call_args = mock_client.capture.call_args[1]
+    call_args = next(
+        c[1]
+        for c in mock_client.capture.call_args_list
+        if c[1]["event"] == "$ai_generation"
+    )
     props = call_args["properties"]
 
     assert call_args["distinct_id"] == "test-id"
diff --git a/posthog/test/ai/openai/test_openai.py b/posthog/test/ai/openai/test_openai.py
index e82bb902..014686a8 100644
--- a/posthog/test/ai/openai/test_openai.py
+++ b/posthog/test/ai/openai/test_openai.py
@@ -739,9 +739,13 @@ def test_tool_calls(mock_client, mock_openai_response_with_tool_calls):
         )
 
         assert response == mock_openai_response_with_tool_calls
-        assert mock_client.capture.call_count == 1
+        # One $ai_generation plus one $ai_span for the tool call the model requested
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        calls = mock_client.capture.call_args_list
+        generation_call = next(c[1] for c in calls if c[1]["event"] == "$ai_generation")
+        span_call = next(c[1] for c in calls if c[1]["event"] == "$ai_span")
+        call_args = generation_call
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -786,6 +790,20 @@ def test_tool_calls(mock_client, mock_openai_response_with_tool_calls):
         assert props["$ai_output_tokens"] == 15
         assert props["$ai_http_status"] == 200
 
+        # The tool call the model requested is emitted as a child $ai_span,
+        # nested under the generation.
+        span_props = span_call["properties"]
+        assert span_call["distinct_id"] == "test-id"
+        assert span_props["$ai_span_type"] == "tool"
+        assert span_props["$ai_span_name"] == "get_weather"
+        assert span_props["$ai_tool_call_id"] == "call_abc123"
+        assert (
+            span_props["$ai_input_state"]
+            == '{"location": "San Francisco", "unit": "celsius"}'
+        )
+        assert span_props["$ai_trace_id"] == props["$ai_trace_id"]
+        assert span_props["$ai_parent_id"] == props["$ai_span_id"]
+
 
 def test_tool_calls_only_no_content(mock_client, mock_openai_response_tool_calls_only):
     with patch(
@@ -810,9 +828,13 @@ def test_tool_calls_only_no_content(mock_client, mock_openai_response_tool_calls
         )
 
         assert response == mock_openai_response_tool_calls_only
-        assert mock_client.capture.call_count == 1
+        # One $ai_generation plus one $ai_span for the requested tool call
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        calls = mock_client.capture.call_args_list
+        generation_call = next(c[1] for c in calls if c[1]["event"] == "$ai_generation")
+        span_call = next(c[1] for c in calls if c[1]["event"] == "$ai_span")
+        call_args = generation_call
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -840,6 +862,14 @@ def test_tool_calls_only_no_content(mock_client, mock_openai_response_tool_calls
         assert props["$ai_output_tokens"] == 10
         assert props["$ai_http_status"] == 200
 
+        # Tool call emitted as a child $ai_span nested under the generation
+        span_props = span_call["properties"]
+        assert span_props["$ai_span_type"] == "tool"
+        assert span_props["$ai_span_name"] == "get_weather"
+        assert span_props["$ai_tool_call_id"] == "call_def456"
+        assert span_props["$ai_input_state"] == '{"location": "New York"}'
+        assert span_props["$ai_parent_id"] == props["$ai_span_id"]
+
 
 def test_responses_api_tool_calls(mock_client, mock_responses_api_with_tool_calls):
     with patch(
@@ -865,9 +895,13 @@ def test_responses_api_tool_calls(mock_client, mock_responses_api_with_tool_call
         )
 
         assert response == mock_responses_api_with_tool_calls
-        assert mock_client.capture.call_count == 1
+        # One $ai_generation plus one $ai_span for the requested tool call
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        calls = mock_client.capture.call_args_list
+        generation_call = next(c[1] for c in calls if c[1]["event"] == "$ai_generation")
+        span_call = next(c[1] for c in calls if c[1]["event"] == "$ai_span")
+        call_args = generation_call
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -897,6 +931,14 @@ def test_responses_api_tool_calls(mock_client, mock_responses_api_with_tool_call
         assert props["$ai_output_tokens"] == 20
         assert props["$ai_http_status"] == 200
 
+        # Tool call emitted as a child $ai_span nested under the generation
+        span_props = span_call["properties"]
+        assert span_props["$ai_span_type"] == "tool"
+        assert span_props["$ai_span_name"] == "get_weather"
+        assert span_props["$ai_tool_call_id"] == "call_xyz789"
+        assert span_props["$ai_input_state"] == '{"location": "Chicago"}'
+        assert span_props["$ai_parent_id"] == props["$ai_span_id"]
+
 
 def test_streaming_with_tool_calls(mock_client, streaming_tool_call_chunks):
     # Mock the create method to return our chunks
@@ -933,10 +975,14 @@ def test_streaming_with_tool_calls(mock_client, streaming_tool_call_chunks):
         assert len(chunks) == 4
         assert chunks == streaming_tool_call_chunks
 
-        # Verify the capture was called with the right arguments
-        assert mock_client.capture.call_count == 1
+        # Verify the capture was called with the right arguments:
+        # one $ai_generation plus one $ai_span for the streamed tool call
+        assert mock_client.capture.call_count == 2
 
-        call_args = mock_client.capture.call_args[1]
+        calls = mock_client.capture.call_args_list
+        generation_call = next(c[1] for c in calls if c[1]["event"] == "$ai_generation")
+        span_call = next(c[1] for c in calls if c[1]["event"] == "$ai_span")
+        call_args = generation_call
         props = call_args["properties"]
 
         assert call_args["distinct_id"] == "test-id"
@@ -995,6 +1041,17 @@ def test_streaming_with_tool_calls(mock_client, streaming_tool_call_chunks):
         assert "prompt_tokens" in props["$ai_usage"]
         assert "completion_tokens" in props["$ai_usage"]
 
+        # The accumulated tool call is emitted as a child $ai_span
+        span_props = span_call["properties"]
+        assert span_props["$ai_span_type"] == "tool"
+        assert span_props["$ai_span_name"] == "get_weather"
+        assert span_props["$ai_tool_call_id"] == "call_abc123"
+        assert (
+            span_props["$ai_input_state"]
+            == '{"location": "San Francisco", "unit": "celsius"}'
+        )
+        assert span_props["$ai_parent_id"] == props["$ai_span_id"]
+
 
 # test responses api
 def test_responses_api(mock_client, mock_openai_response_with_responses_api):