fix(main): apply a call's flat rate to its own tiered_pricing table when it supplies one

litellm/main.py

@@ -1502,9 +1502,10 @@ def _token_rates_cleared_for_per_second_overrides(custom: Mapping[str, object])
 
 def _tiers_with_flat_overrides(catalog: Mapping[str, object], custom: Mapping[str, object]) -> Mapping[str, object]:
     """A ``tiered_pricing`` table wins over flat rates in the cost calculator, so a flat rate the caller set is written
-    into every tier for that direction, and the other direction keeps each tier's own rate."""
+    into every tier of the table in force (the caller's own, else the inherited one) for that direction, and the other
+    direction keeps each tier's own rate."""
     overrides: Final = {key: custom[key] for key in _FLAT_TOKEN_RATE_FIELDS if key in custom}
-    tiers: Final = catalog.get("tiered_pricing")
+    tiers: Final = custom.get("tiered_pricing", catalog.get("tiered_pricing"))
     if not overrides or not isinstance(tiers, (list, tuple)) or not tiers:
         return {}
     return {"tiered_pricing": [{**tier, **overrides} for tier in _TIERS.validate_python(tiers)]}

tests/unit/test_register_model_custom_pricing.py

@@ -1667,3 +1667,35 @@ def test_direct_call_with_only_a_per_second_rate_on_a_mapped_model_is_billed_per
     finally:
         litellm.model_cost.pop(model_key, None)
         _invalidate_model_cost_lowercase_map()
+
+
+def test_direct_call_flat_rate_applies_to_the_callers_own_tiers_not_the_catalogs():
+    model_key: Final = "openai/lit-caller-tiers-model"
+    litellm.register_model(
+        {
+            model_key: {
+                "litellm_provider": "openai",
+                "mode": "chat",
+                "tiered_pricing": [
+                    {"range": [0, 1000000], "input_cost_per_token": 3e-07, "output_cost_per_token": 6e-07}
+                ],
+            }
+        },
+        persist_across_reloads=False,
+    )
+    try:
+        response: Final = litellm.completion(
+            model=model_key,
+            messages=[{"role": "user", "content": "hello"}],
+            api_key="fake-key",
+            mock_response="ok",
+            input_cost_per_token=1e-05,
+            tiered_pricing=[{"range": [0, 1000000], "input_cost_per_token": 5e-06, "output_cost_per_token": 9e-06}],
+        )
+        assert response._hidden_params["response_cost"] == pytest.approx(
+            response.usage.prompt_tokens * 1e-05 + response.usage.completion_tokens * 9e-06
+        )
+        assert litellm.model_cost[model_key]["tiered_pricing"][0]["output_cost_per_token"] == 6e-07
+    finally:
+        litellm.model_cost.pop(model_key, None)
+        _invalidate_model_cost_lowercase_map()