fix(decisions): keep a token-priced row on token billing when it also sets a query price

A plain input_cost_per_query bills flat only on a row with no token price, and input_cost_per_query_decisions always bills flat, so a custom Decisions row that already bills tokens costs the same as before this PR
This commit is contained in:
mateo-berri 2026-10-31 13:21:06 -07:00
parent 54dac276d3
commit 56039b9207
4 changed files with 47 additions and 13 deletions

View file

@ -24,7 +24,11 @@ from litellm.constants import (
)
from litellm.litellm_core_utils.core_helpers import as_str_mapping, billed_usage_cost
from litellm.litellm_core_utils.cost_discount import resolve_cost_discount
from litellm.litellm_core_utils.decisions_pricing import decisions_pricing_fields, is_decisions_call_type
from litellm.litellm_core_utils.decisions_pricing import (
decisions_pricing_fields,
decisions_request_price,
is_decisions_call_type,
)
from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR
from litellm.litellm_core_utils.llm_cost_calc.tool_call_cost_tracking import (
StandardBuiltInToolCostTracking,
@ -604,11 +608,9 @@ def cost_per_token(
decisions_overrides: Final = decisions_pricing_fields(
litellm.model_cost.get(decisions_model_info["key"], {}) # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType] # model cost map is dynamically loaded
)
request_price: Final = decisions_overrides.get(
"input_cost_per_query", decisions_model_info.get("input_cost_per_query")
)
if isinstance(request_price, (int, float)):
return float(request_price), 0.0
request_price: Final = decisions_request_price(decisions_model_info, decisions_overrides)
if request_price is not None:
return request_price, 0.0
if decisions_overrides:
return generic_cost_per_token(
model=model,

View file

@ -27,6 +27,15 @@ def decisions_pricing_fields(row: Mapping[str, object]) -> Mapping[str, object]:
}
def decisions_request_price(model_info: Mapping[str, object], overrides: Mapping[str, object]) -> float | None:
rates: Final = {**model_info, **overrides}
request_price: Final = rates.get("input_cost_per_query")
if not isinstance(request_price, (int, float)):
return None
bills_tokens: Final = any(rates.get(field) for field in ("input_cost_per_token", "output_cost_per_token"))
return None if bills_tokens and "input_cost_per_query" not in overrides else float(request_price)
def apply_decisions_pricing(row: Mapping[str, object]) -> Mapping[str, object]:
"""Return ``row`` priced for a Decisions call: ``_decisions`` values replace their plain fields."""
plain: Final = {key: value for key, value in row.items() if not key.endswith(DECISIONS_PRICING_SUFFIX)}

View file

@ -1,10 +1,3 @@
"""Parallel's decision model uses System One questions and the native fast mode.
Both LiteLLM Decisions routes share this adapter. Citations describe the whole
request, so they stay on the response rather than being assigned to answers.
Contract: https://docs.parallel.ai/api-reference/decisions/make-web-grounded-typed-decisions
"""
from collections.abc import Mapping
from typing import Final

View file

@ -77,6 +77,36 @@ def test_flat_decisions_rate_overrides_catalog_without_changing_chat(
) == pytest.approx((100 * 1e-07, 0.0))
@pytest.mark.parametrize("call_type", ["adecisions", "asystemone"])
@pytest.mark.parametrize(
("token_rates", "expected_costs"),
[
({"input_cost_per_token": 1e-05, "output_cost_per_token": 0.0}, (100 * 1e-05, 0.0)),
({"input_cost_per_token": 0.0, "output_cost_per_token": 1e-05}, (0.0, 10 * 1e-05)),
],
)
def test_plain_query_price_leaves_a_token_priced_row_billing_tokens(
monkeypatch: pytest.MonkeyPatch,
call_type: str,
token_rates: dict[str, float],
expected_costs: tuple[float, float],
) -> None:
model: Final = "fireworks_ai/accounts/example/models/priced-both"
monkeypatch.setitem(
litellm.model_cost,
model,
{"input_cost_per_query": 0.5, **token_rates, "litellm_provider": "fireworks_ai", "mode": "evaluation"},
)
assert cost_calculator.cost_per_token(
model=model,
custom_llm_provider="fireworks_ai",
prompt_tokens=100,
completion_tokens=10,
call_type=cast(CallTypesLiteral, call_type),
) == pytest.approx(expected_costs)
_LUNA_USAGE: Final = Usage(
prompt_tokens=100,
completion_tokens=10,