From 91808ed2cf53dbeed76ed63908cea8b9cbcaf24f Mon Sep 17 00:00:00 2001
From: Kerry Lu <kerry@berri.ai>
Date: Thu, 12 Nov 2026 15:34:24 -0800
Subject: [PATCH] fix(fal_ai): bill Luma Ray v3.2 videos at fal's 540p default
 resolution (#51477)

fal's luma/agent/ray/v3.2 endpoints default resolution to 540p, but the fal video transformation had no
Luma profile and fell back to the Seedance profile, whose default is 720p. A request that left
resolution unset was stamped video_resolution 720p and priced from output_cost_per_second_720p, which is
twice the 540p rate on text-to-video, image-to-video and reframe and 1.5x on video-to-video.

A Luma Ray profile (540p, 720p, 1080p, default 540p) stamps the real default, so those requests price
from output_cost_per_second, the 540p rate.

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/llms/fal_ai/videos/transformation.py  | 10 ++++++++
 .../test_fal_ai_video_transformation.py       | 23 +++++++++++++++++++
 2 files changed, 33 insertions(+)

diff --git a/litellm/llms/fal_ai/videos/transformation.py b/litellm/llms/fal_ai/videos/transformation.py
index 838e7065ef7..d6cd7f0d5ce 100644
--- a/litellm/llms/fal_ai/videos/transformation.py
+++ b/litellm/llms/fal_ai/videos/transformation.py
@@ -75,6 +75,14 @@ _H3_MAX_PROFILE: Final[_ModelProfile] = _ModelProfile(
     reference_key="image_url",
     reference_as_list=False,
 )
+_LUMA_RAY_PROFILE: Final[_ModelProfile] = _ModelProfile(
+    resolutions=frozenset({"540p", "720p", "1080p"}),
+    resolution_tiers=((540, "540p"), (720, "720p"), (sys.maxsize, "1080p")),
+    default_resolution="540p",
+    integer_duration=False,
+    reference_key="image_url",
+    reference_as_list=False,
+)
 _SEEDANCE_REFERENCE_PROFILE: Final[_ModelProfile] = replace(
     _SEEDANCE_PROFILE,
     reference_key="image_urls",
@@ -118,6 +126,8 @@ def _profile_for_model(model: str) -> _ModelProfile:
         return _H3_PROFILE
     if model.startswith("minimax/h3-max"):
         return _H3_MAX_PROFILE
+    if model.startswith("luma/agent/ray/"):
+        return _LUMA_RAY_PROFILE
     is_seedance_reference: Final = model.startswith("bytedance/seedance-") and model.endswith("/reference-to-video")
     return _SEEDANCE_REFERENCE_PROFILE if is_seedance_reference else _SEEDANCE_PROFILE
 
diff --git a/tests/unit/llms/fal_ai/videos/test_fal_ai_video_transformation.py b/tests/unit/llms/fal_ai/videos/test_fal_ai_video_transformation.py
index b20c8069789..3498663775f 100644
--- a/tests/unit/llms/fal_ai/videos/test_fal_ai_video_transformation.py
+++ b/tests/unit/llms/fal_ai/videos/test_fal_ai_video_transformation.py
@@ -232,6 +232,29 @@ class TestFalAIVideoTransformation:
 
         assert video.usage == {"duration_seconds": 5.0, "video_resolution": "2K"}
 
+    def test_transform_video_create_response_luma_ray_defaults_to_540p(self):
+        response = Mock(spec=httpx.Response)
+        response.json.return_value = {"request_id": "abc"}
+
+        for model in (
+            "luma/agent/ray/v3.2/text-to-video",
+            "luma/agent/ray/v3.2/image-to-video",
+            "luma/agent/ray/v3.2/reframe",
+            "luma/agent/ray/v3.2/video-to-video",
+        ):
+            video = self.config.transform_video_create_response(
+                model=model,
+                raw_response=response,
+                logging_obj=self.logging_obj,
+                custom_llm_provider="fal_ai",
+                request_data={"duration": "5"},
+            )
+            assert video.usage == {"duration_seconds": 5.0, "video_resolution": "540p"}, model
+            row = litellm.model_cost[f"fal_ai/{model}"]
+            assert default_video_cost_calculator(model, 5, "fal_ai", video_resolution="540p") == (
+                5 * row["output_cost_per_second"]
+            ), model
+
     def test_transform_video_create_response_h3_max_leaves_default_resolution_to_cost_map(
         self, local_model_cost_map: None
     ) -> None: