feat(ai): emit the served service tier as the explicit $ai_service_tier property (#932)

* feat(ai): emit the served tier as the explicit $ai_service_tier property

Cost processing prices only from this property: its writers assert
served values, unlike $ai_model_parameters.service_tier, which older
releases populated from the request.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: add changeset

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: bernatixer

.sampo/changesets/emit-explicit-served-tier.md

@@ -0,0 +1,5 @@
+---
+pypi/posthog: minor
+---
+
+OpenAI and LangChain generations now also emit the served service tier as the explicit `$ai_service_tier` event property, next to the copy inside `$ai_model_parameters`. Cost processing prices tiered calls only from the explicit property, whose writers assert response-derived values.

posthog/ai/langchain/callbacks.py

@@ -619,6 +619,7 @@ def _capture_generation(
     ):
         # The served tier comes from the response, because a requested tier can be refused.
         model_params = run.model_params
+        served_tier = None
         if isinstance(output, LLMResult) and isinstance(output.llm_output, dict):
             served_tier = output.llm_output.get("service_tier")
             if served_tier is not None:
@@ -632,6 +633,8 @@ def _capture_generation(
             "$ai_provider": run.provider,
             "$ai_model": run.model,
             "$ai_model_parameters": model_params,
+            # The explicit served-tier signal cost processing prices from.
+            **({"$ai_service_tier": served_tier} if served_tier is not None else {}),
             "$ai_input": with_privacy_mode(
                 self._ph_client,
                 self._privacy_mode,

posthog/ai/utils.py

@@ -521,10 +521,11 @@ def call_llm_and_track_usage(
 
             tag("$ai_provider", provider)
             tag("$ai_model", kwargs.get("model") or getattr(response, "model", None))
-            tag(
-                "$ai_model_parameters",
-                get_model_params(kwargs, getattr(response, "service_tier", None)),
-            )
+            served_service_tier = getattr(response, "service_tier", None)
+            tag("$ai_model_parameters", get_model_params(kwargs, served_service_tier))
+            if served_service_tier is not None:
+                # The explicit served-tier signal cost processing prices from.
+                tag("$ai_service_tier", served_service_tier)
             tag(
                 "$ai_input",
                 with_privacy_mode(ph_client, posthog_privacy_mode, sanitized_messages),
@@ -687,10 +688,11 @@ async def call_llm_and_track_usage_async(
 
             tag("$ai_provider", provider)
             tag("$ai_model", kwargs.get("model") or getattr(response, "model", None))
-            tag(
-                "$ai_model_parameters",
-                get_model_params(kwargs, getattr(response, "service_tier", None)),
-            )
+            served_service_tier = getattr(response, "service_tier", None)
+            tag("$ai_model_parameters", get_model_params(kwargs, served_service_tier))
+            if served_service_tier is not None:
+                # The explicit served-tier signal cost processing prices from.
+                tag("$ai_service_tier", served_service_tier)
             tag(
                 "$ai_input",
                 with_privacy_mode(ph_client, posthog_privacy_mode, sanitized_messages),
@@ -840,6 +842,11 @@ def capture_streaming_event(
         "$ai_model_parameters": get_model_params(
             event_data["kwargs"], event_data.get("service_tier")
         ),
+        **(
+            {"$ai_service_tier": event_data["service_tier"]}
+            if event_data.get("service_tier") is not None
+            else {}
+        ),
         "$ai_input": with_privacy_mode(
             ph_client,
             event_data["privacy_mode"],

posthog/test/ai/langchain/test_callbacks.py

@@ -2864,6 +2864,7 @@ def test_served_service_tier_merges_into_model_parameters(mock_client):
     props = mock_client.capture.call_args.kwargs["properties"]
     assert props["$ai_model_parameters"]["service_tier"] == "flex"
     assert props["$ai_model_parameters"]["temperature"] == 0.5
+    assert props["$ai_service_tier"] == "flex"
 
 
 @pytest.mark.parametrize(

posthog/test/ai/openai/test_openai.py

@@ -2683,6 +2683,7 @@ def test_served_service_tier_lands_in_model_parameters(
 
         props = mock_client.capture.call_args[1]["properties"]
         assert props["$ai_model_parameters"]["service_tier"] == "flex"
+        assert props["$ai_service_tier"] == "flex"
 
 
 def test_response_without_service_tier_omits_it(mock_client, mock_openai_response):
@@ -2699,6 +2700,7 @@ def test_response_without_service_tier_omits_it(mock_client, mock_openai_respons
 
         props = mock_client.capture.call_args[1]["properties"]
         assert "service_tier" not in props["$ai_model_parameters"]
+        assert "$ai_service_tier" not in props
 
 
 def test_streaming_state_tracks_served_service_tier():

posthog/test/test_utils.py

@@ -112,8 +112,8 @@ def test_guess_timezone_preserves_recent_naive_local_instant_as_utc(self):
             utc_now = datetime(2026, 1, 15, 12, 30, 45, 123456, tzinfo=timezone.utc)
 
             with mock.patch("posthog.utils.datetime", wraps=datetime) as mock_datetime:
-                mock_datetime.now.side_effect = (
-                    lambda tz=None: local_now if tz is None else utc_now.astimezone(tz)
+                mock_datetime.now.side_effect = lambda tz=None: (
+                    local_now if tz is None else utc_now.astimezone(tz)
                 )
                 normalized = utils.guess_timezone(local_now)