From 9fc77f12227e36a6d8e86336e1995931659f1c25 Mon Sep 17 00:00:00 2001 From: Srivatsa03 Date: Tue, 30 Jun 2026 12:10:48 -0500 Subject: [PATCH] feat(cost): support time-based off-peak pricing in cost calculation Some providers charge different per-token rates depending on the time of day. DeepSeek, for example, has historically discounted its chat and reasoner models during an off-peak window (16:30-00:30 UTC). LiteLLM's cost map only modeled static per-token pricing, so cost tracking could not stay accurate for these providers. This adds optional off-peak pricing to a model entry: input_cost_per_token_off_peak, output_cost_per_token_off_peak, cache_read_input_token_cost_off_peak, and an off_peak_hours_utc window expressed as "HH:MM-HH:MM" in UTC (the window may wrap past midnight). When the current UTC time falls inside the window, the cost calculator uses the off-peak rates and otherwise falls back to the standard rates, so existing models are unaffected. The fields are also accepted as custom pricing on a deployment, so they can be set from the proxy config or the SDK. The window check is a pure function that takes the current time as an argument, which keeps the regression tests deterministic without patching the clock. --- .../litellm_core_utils/llm_cost_calc/utils.py | 72 +++++++++ litellm/types/utils.py | 14 ++ litellm/utils.py | 1 + .../llm_cost_calc/test_llm_cost_calc_utils.py | 152 ++++++++++++++++++ 4 files changed, 239 insertions(+) diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py index 19e3f624268..576efb18bb0 100644 --- a/litellm/litellm_core_utils/llm_cost_calc/utils.py +++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py @@ -4,6 +4,7 @@ import re from collections.abc import Mapping from dataclasses import dataclass +from datetime import datetime, timezone from types import MappingProxyType from typing import Any, Final, Literal, TypedDict, cast @@ -290,10 +291,75 @@ def _get_tiered_base_costs(model_info: ModelInfo, usage: Usage) -> tuple[float, ) +def _is_within_off_peak_window(off_peak_hours_utc: str | list[str], current_time: datetime | None = None) -> bool: + """Return True if current_time (UTC, defaulting to now) falls inside any off-peak window. + + off_peak_hours_utc is a "HH:MM-HH:MM" string in UTC, or a list of such strings for providers + with multiple daily windows (e.g. ["16:30-00:30", "04:00-06:00"]). A window may wrap past + midnight. The start is inclusive and the end is exclusive; malformed windows are ignored. + """ + if current_time is None: + current_time = datetime.now(timezone.utc) + now = current_time.time() + windows = [off_peak_hours_utc] if isinstance(off_peak_hours_utc, str) else off_peak_hours_utc + for window in windows: + try: + start_str, end_str = window.split("-") + start = datetime.strptime(start_str.strip(), "%H:%M").time() + end = datetime.strptime(end_str.strip(), "%H:%M").time() + except (ValueError, AttributeError): + continue + if start <= end: + if start <= now < end: + return True + elif now >= start or now < end: + return True + return False + + +def _coerce_off_peak_rate(value: object, default: float) -> float: + if isinstance(value, bool): + return default + if isinstance(value, (int, float)): + return float(value) + if isinstance(value, str): + try: + return float(value) + except ValueError: + return default + return default + + +def _apply_off_peak_pricing( + model_info: ModelInfo, + current_time: datetime | None, + prompt_base_cost: float, + completion_base_cost: float, + cache_read_cost: float, +) -> tuple[float, float, float]: + """Swap in off-peak per-token rates when the current UTC time is inside one of the model's + off_peak_pricing windows. Applied after threshold pricing so the discount is honored rather + than overwritten when a model combines off-peak and above-threshold rates. Any rate left + unset in off_peak_pricing falls back to the standard rate. + """ + off_peak = model_info.get("off_peak_pricing") + if not off_peak: + return prompt_base_cost, completion_base_cost, cache_read_cost + hours_utc = off_peak.get("hours_utc") + if not hours_utc or not _is_within_off_peak_window(hours_utc, current_time): + return prompt_base_cost, completion_base_cost, cache_read_cost + return ( + _coerce_off_peak_rate(off_peak.get("input_cost_per_token"), prompt_base_cost), + _coerce_off_peak_rate(off_peak.get("output_cost_per_token"), completion_base_cost), + _coerce_off_peak_rate(off_peak.get("cache_read_input_token_cost"), cache_read_cost), + ) + + def _get_token_base_cost( model_info: ModelInfo, usage: Usage, service_tier: str | None = None, + current_time: datetime | None = None, *, threshold_is_inclusive: bool = False, ) -> tuple[float, float, float, float, float]: @@ -345,6 +411,9 @@ def _get_token_base_cost( k for k in model_info if k.startswith("input_cost_per_token_above_") and not k.endswith(_SERVICE_TIER_SUFFIXES) ] if not threshold_keys: + prompt_base_cost, completion_base_cost, cache_read_cost = _apply_off_peak_pricing( + model_info, current_time, prompt_base_cost, completion_base_cost, cache_read_cost + ) return ( prompt_base_cost, completion_base_cost, @@ -451,6 +520,9 @@ def _get_token_base_cost( except Exception: continue + prompt_base_cost, completion_base_cost, cache_read_cost = _apply_off_peak_pricing( + model_info, current_time, prompt_base_cost, completion_base_cost, cache_read_cost + ) return ( prompt_base_cost, completion_base_cost, diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 4bf8289d725..6583b125b62 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -193,6 +193,19 @@ class AgenticLoopParams(TypedDict, total=False): """The LLM provider name (e.g., 'bedrock', 'anthropic')""" +class OffPeakPricing(TypedDict, total=False): + """Time-windowed off-peak rates for providers that discount by time of day (e.g. DeepSeek). + + hours_utc is a "HH:MM-HH:MM" string in UTC, or a list of them for multiple daily windows; + a window may wrap past midnight. Any rate left unset falls back to the standard rate. + """ + + hours_utc: str | list[str] + input_cost_per_token: float + output_cost_per_token: float + cache_read_input_token_cost: float + + class ModelInfoBase(ProviderSpecificModelInfo, total=False): key: Required[str] # the key in litellm.model_cost which is returned @@ -225,6 +238,7 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False): # Smallest prefix this model will actually cache, whatever caching mechanism its provider uses. # Absent means the provider-agnostic default applies; see MINIMUM_PROMPT_CACHE_TOKEN_COUNT. prompt_cache_min_tokens: int | None + off_peak_pricing: OffPeakPricing | None # time-windowed off-peak rates input_cost_per_character: float | None # only for vertex ai models input_cost_per_audio_token: float | None input_cost_per_token_above_128k_tokens: float | None # only for vertex ai models diff --git a/litellm/utils.py b/litellm/utils.py index 5e9e115ed54..3389b8fcb78 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -5842,6 +5842,7 @@ def _get_model_info_helper( cache_creation_input_token_cost_above_1hr=_model_info.get( "cache_creation_input_token_cost_above_1hr", None ), + off_peak_pricing=_model_info.get("off_peak_pricing", None), input_cost_per_character=_model_info.get("input_cost_per_character", None), input_cost_per_token_above_128k_tokens=_model_info.get("input_cost_per_token_above_128k_tokens", None), input_cost_per_token_above_200k_tokens=_model_info.get("input_cost_per_token_above_200k_tokens", None), diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py index 3c4121977de..e226e255b05 100644 --- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py +++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py @@ -32,6 +32,7 @@ from litellm.litellm_core_utils.llm_cost_calc.utils import ( TokenTypeCostBreakdown, _calculate_input_cost, _get_token_base_cost, + _is_within_off_peak_window, calculate_cache_writing_cost, generic_cost_per_token, get_token_type_cost_breakdown, @@ -3946,3 +3947,154 @@ def test_route_image_generation_cost_falls_back_to_requested_size(monkeypatch, r ) assert cost == expected_cost + + +def test_is_within_off_peak_window_same_day(): + from datetime import datetime, timezone + + window = "09:00-17:00" + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 12, 0, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 8, 59, tzinfo=timezone.utc)) is False + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 9, 0, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 17, 0, tzinfo=timezone.utc)) is False + + +def test_is_within_off_peak_window_wraps_midnight(): + from datetime import datetime, timezone + + window = "16:30-00:30" + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 0, 15, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 16, 30, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 0, 30, tzinfo=timezone.utc)) is False + assert _is_within_off_peak_window(window, datetime(2026, 1, 1, 12, 0, tzinfo=timezone.utc)) is False + + +def test_is_within_off_peak_window_multiple_windows(): + from datetime import datetime, timezone + + # Providers like DeepSeek V4 have more than one daily peak/off-peak window. + windows = ["01:00-05:00", "13:00-16:00"] + assert _is_within_off_peak_window(windows, datetime(2026, 1, 1, 3, 0, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(windows, datetime(2026, 1, 1, 14, 30, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window(windows, datetime(2026, 1, 1, 9, 0, tzinfo=timezone.utc)) is False + # a malformed entry in the list is ignored, valid entries still match + assert _is_within_off_peak_window(["bad", "13:00-16:00"], datetime(2026, 1, 1, 14, 0, tzinfo=timezone.utc)) is True + assert _is_within_off_peak_window([], datetime(2026, 1, 1, 14, 0, tzinfo=timezone.utc)) is False + + +def test_is_within_off_peak_window_malformed_returns_false(): + from datetime import datetime, timezone + + now = datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc) + assert _is_within_off_peak_window("not-a-window", now) is False + assert _is_within_off_peak_window("16:30", now) is False + assert _is_within_off_peak_window("25:00-26:00", now) is False + + +def test_get_token_base_cost_applies_off_peak_pricing(): + from datetime import datetime, timezone + from typing import cast + + from litellm.types.utils import ModelInfo + + model_info = cast( + ModelInfo, + { + "input_cost_per_token": 1e-6, + "output_cost_per_token": 2e-6, + "cache_read_input_token_cost": 1e-7, + "off_peak_pricing": { + "hours_utc": "16:30-00:30", + "input_cost_per_token": 5e-7, + "output_cost_per_token": 1e-6, + "cache_read_input_token_cost": 5e-8, + }, + }, + ) + usage = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150) + + off_peak = _get_token_base_cost(model_info, usage, current_time=datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc)) + assert off_peak[0] == 5e-7 + assert off_peak[1] == 1e-6 + assert off_peak[4] == 5e-8 + + peak = _get_token_base_cost(model_info, usage, current_time=datetime(2026, 1, 1, 12, 0, tzinfo=timezone.utc)) + assert peak[0] == 1e-6 + assert peak[1] == 2e-6 + assert peak[4] == 1e-7 + + +def test_get_token_base_cost_off_peak_falls_back_to_standard_when_unset(): + from datetime import datetime, timezone + from typing import cast + + from litellm.types.utils import ModelInfo + + model_info = cast( + ModelInfo, + { + "input_cost_per_token": 1e-6, + "output_cost_per_token": 2e-6, + "off_peak_pricing": {"hours_utc": "16:30-00:30", "input_cost_per_token": 5e-7}, + }, + ) + usage = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150) + + result = _get_token_base_cost(model_info, usage, current_time=datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc)) + assert result[0] == 5e-7 + assert result[1] == 2e-6 + + +def test_get_token_base_cost_off_peak_wins_over_threshold(): + from datetime import datetime, timezone + from typing import cast + + from litellm.types.utils import ModelInfo + + model_info = cast( + ModelInfo, + { + "input_cost_per_token": 1e-6, + "output_cost_per_token": 2e-6, + "input_cost_per_token_above_200k_tokens": 3e-6, + "output_cost_per_token_above_200k_tokens": 4e-6, + "off_peak_pricing": { + "hours_utc": "16:30-00:30", + "input_cost_per_token": 5e-7, + "output_cost_per_token": 1e-6, + }, + }, + ) + usage = Usage(prompt_tokens=250000, completion_tokens=250000, total_tokens=500000) + + off_peak = _get_token_base_cost(model_info, usage, current_time=datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc)) + assert off_peak[0] == 5e-7 + assert off_peak[1] == 1e-6 + + peak = _get_token_base_cost(model_info, usage, current_time=datetime(2026, 1, 1, 12, 0, tzinfo=timezone.utc)) + assert peak[0] == 3e-6 + assert peak[1] == 4e-6 + + +def test_get_model_info_propagates_off_peak_fields(): + model_name = "test-off-peak-model" + off_peak_pricing = { + "hours_utc": "16:30-00:30", + "input_cost_per_token": 5e-7, + "output_cost_per_token": 1e-6, + "cache_read_input_token_cost": 5e-8, + } + litellm.register_model( + { + model_name: { + "litellm_provider": "openai", + "mode": "chat", + "input_cost_per_token": 1e-6, + "output_cost_per_token": 2e-6, + "off_peak_pricing": off_peak_pricing, + } + } + ) + info = litellm.get_model_info(model=model_name) + assert info["off_peak_pricing"] == off_peak_pricing