diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py index 66f47affb4d..be289f620cf 100644 --- a/litellm/litellm_core_utils/llm_cost_calc/utils.py +++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py @@ -337,9 +337,10 @@ def _apply_off_peak_pricing( 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_pricing windows. An off-peak rate replaces the rate that would otherwise apply + rather than discounting it, so a model that also has tiered or above-threshold pricing bills + the flat off-peak rate for the whole request while the window is open. Any rate left unset in + off_peak_pricing falls back to the standard rate. """ off_peak: Final = model_info.get("off_peak_pricing") if not off_peak: @@ -354,6 +355,22 @@ def _apply_off_peak_pricing( ) +def _apply_off_peak_to_base_costs( + model_info: ModelInfo, + current_time: datetime | None, + base_costs: tuple[float, float, float, float, float], +) -> tuple[float, float, float, float, float]: + """Apply off-peak rates to an already-resolved set of base costs, whichever pricing path + produced them. Cache-creation rates are passed through untouched, since off_peak_pricing + has no field for them. + """ + prompt, completion, cache_creation, cache_creation_above_1hr, cache_read = base_costs + off_peak_prompt, off_peak_completion, off_peak_cache_read = _apply_off_peak_pricing( + model_info, current_time, prompt, completion, cache_read + ) + return (off_peak_prompt, off_peak_completion, cache_creation, cache_creation_above_1hr, off_peak_cache_read) + + def _get_token_base_cost( model_info: ModelInfo, usage: Usage, @@ -376,7 +393,7 @@ def _get_token_base_cost( """ tiered_base_costs: Final = _get_tiered_base_costs(model_info=model_info, usage=usage) if tiered_base_costs is not None: - return tiered_base_costs + return _apply_off_peak_to_base_costs(model_info, current_time, tiered_base_costs) # Get service tier aware cost keys input_cost_key: Final = _get_service_tier_cost_key("input_cost_per_token", service_tier) @@ -410,15 +427,16 @@ 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: - off_peak_prompt_cost, off_peak_completion_cost, off_peak_cache_read_cost = _apply_off_peak_pricing( - model_info, current_time, prompt_base_cost, completion_base_cost, cache_read_cost - ) - return ( - off_peak_prompt_cost, - off_peak_completion_cost, - cache_creation_cost, - cache_creation_cost_above_1hr, - off_peak_cache_read_cost, + return _apply_off_peak_to_base_costs( + model_info, + current_time, + ( + prompt_base_cost, + completion_base_cost, + cache_creation_cost, + cache_creation_cost_above_1hr, + cache_read_cost, + ), ) # Only sort the threshold keys (typically 1-2 keys instead of 66+) @@ -519,15 +537,16 @@ def _get_token_base_cost( except Exception: continue - discounted_prompt_cost, discounted_completion_cost, discounted_cache_read_cost = _apply_off_peak_pricing( - model_info, current_time, prompt_base_cost, completion_base_cost, cache_read_cost - ) - return ( - discounted_prompt_cost, - discounted_completion_cost, - cache_creation_cost, - cache_creation_cost_above_1hr, - discounted_cache_read_cost, + return _apply_off_peak_to_base_costs( + model_info, + current_time, + ( + prompt_base_cost, + completion_base_cost, + cache_creation_cost, + cache_creation_cost_above_1hr, + cache_read_cost, + ), ) 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 f1340df8b69..8f348b36d35 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 @@ -574,6 +574,38 @@ def test_get_model_info_propagates_off_peak_fields(): assert info["off_peak_pricing"] == off_peak_pricing +def test_get_token_base_cost_off_peak_wins_over_tiered_pricing(): + """Tiered pricing resolves base rates on its own path and returns early, so off-peak has to + be applied there too or a model carrying both would silently bill the tier rate all day.""" + from datetime import datetime, timezone + + model_name = "litellm-test-off-peak-tiered" + litellm.register_model( + { + model_name: { + "litellm_provider": "openai", + "mode": "chat", + "tiered_pricing": [ + {"range": [0, 128000], "input_cost_per_token": 3e-6, "output_cost_per_token": 6e-6}, + ], + "off_peak_pricing": { + "hours_utc": "16:30-00:30", + "input_cost_per_token": 5e-7, + "output_cost_per_token": 1e-6, + }, + } + } + ) + info = litellm.get_model_info(model=model_name) + usage = Usage(prompt_tokens=1_000, completion_tokens=100, total_tokens=1_100) + + inside = _get_token_base_cost(info, usage, current_time=datetime(2026, 1, 1, 18, 0, tzinfo=timezone.utc)) + assert inside[:2] == (5e-7, 1e-6) + + outside = _get_token_base_cost(info, usage, current_time=datetime(2026, 1, 1, 12, 0, tzinfo=timezone.utc)) + assert outside[:2] == (3e-6, 6e-6) + + def test_generic_cost_per_token_gpt54_above_272k_tokens(_local_model_cost_map): """GPT-5.4/5.4-pro: prompts >272K input tokens priced at 2x input, 1.5x output.""" model = "gpt-5.4"