From fcb22cb20226d2394e980096ba2e8d30c0e29b5c Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 28 Oct 2026 05:53:52 -0700 Subject: [PATCH] fix(cost): bill the requested model's rates when the response names an unpriced model Fireworks answers chat and /v1/messages calls with a display name such as "Qwen 3.8 Max" in the response's model field. The cost path priced that name, the Fireworks calculator fell back to its free default bucket without raising, and the call billed $0 even though the requested model has a row. When the response model resolves to no cost-map row or rule that declares rates and the requested model does, price from the requested model. A priced response model still wins, so dated snapshots keep their own rates, and an unpriced requested model such as a router alias leaves the result unchanged. --- litellm/cost_calculator.py | 49 ++++++++------ tests/unit/test_cost_calculator.py | 100 +++++++++++++++++++++++++++++ 2 files changed, 130 insertions(+), 19 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 86b22069a39..bc7dbff47f9 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -961,19 +961,14 @@ def select_model_name_for_cost_calc( """ 1. If custom pricing is true, return received model name 2. If base_model is set (e.g. for azure models), return that - 3. If completion response has model set return that + 3. If completion response has model set return that, unless only the requested model's cost-map row has rates 4. Check if model is passed in return that """ return_model: str | None = None custom_llm_provider = _get_provider_for_cost_calc(model=model, custom_llm_provider=custom_llm_provider) - completion_response_model: str | None = None - if completion_response is not None: - if isinstance(completion_response, BaseModel): - completion_response_model = getattr(completion_response, "model", None) - elif isinstance(completion_response, dict): - completion_response_model = completion_response.get("model", None) + completion_response_model: Final = _get_response_model(completion_response) hidden_params: Final[dict | None] = getattr(completion_response, "_hidden_params", None) provider_response_model: Final = _get_hidden_str_for_cost_calc(hidden_params, "provider_response_model") explicit_pricing: Final = custom_pricing is True or base_model is not None @@ -1009,27 +1004,43 @@ def select_model_name_for_cost_calc( return_model = hidden_params.get("model", model) if return_model is None and completion_response_model is not None: - return_model = completion_response_model + return _served_or_priced_requested_model( + served=_cost_map_model_name(completion_response_model, custom_llm_provider, priced_region), + requested=_cost_map_model_name(model, custom_llm_provider, None) if model is not None else None, + custom_llm_provider=custom_llm_provider, + ) if return_model is None and model is not None: return_model = model - if ( - return_model is not None - and custom_llm_provider is not None - and not _model_contains_known_llm_provider(return_model) - ): # add provider prefix if not already present, to match model_cost - provider_prefix: Final = ( - custom_llm_provider if priced_region is None else f"{custom_llm_provider}/{priced_region}" - ) - return_model = _strip_unregistered_leading_segments(f"{provider_prefix}/{return_model}", priced_region) - - return return_model + if return_model is None: + return None + return _cost_map_model_name(return_model, custom_llm_provider, priced_region) _select_model_name_for_cost_calc = select_model_name_for_cost_calc +def _cost_map_model_name(model: str, custom_llm_provider: str | None, region_name: str | None) -> str: + if custom_llm_provider is None or _model_contains_known_llm_provider(model): + return model + provider_prefix: Final = custom_llm_provider if region_name is None else f"{custom_llm_provider}/{region_name}" + return _strip_unregistered_leading_segments(f"{provider_prefix}/{model}", region_name) + + +def _resolves_to_rates(model: str, custom_llm_provider: str | None) -> bool: + rule_info: Final = unmapped_model_generalization(model, custom_llm_provider) + if rule_info is not None: + return _cost_map_entry_prices_anything(rule_info) + return _declares_rates(_cost_map_model_info(model, custom_llm_provider)) + + +def _served_or_priced_requested_model(served: str, requested: str | None, custom_llm_provider: str | None) -> str: + if requested is None or requested == served or _resolves_to_rates(served, custom_llm_provider): + return served + return requested if _resolves_to_rates(requested, custom_llm_provider) else served + + def _strip_unregistered_leading_segments(model: str, region_name: str | None) -> str: """Resolve a provider-prefixed slash alias like "vertex_ai/vertex/claude-opus-5" to the registered cost key ("vertex_ai/claude-opus-5"), keeping the model unchanged when it already diff --git a/tests/unit/test_cost_calculator.py b/tests/unit/test_cost_calculator.py index 84f6dab9eff..5886e0ce2ef 100644 --- a/tests/unit/test_cost_calculator.py +++ b/tests/unit/test_cost_calculator.py @@ -424,6 +424,106 @@ def test_completion_cost_prices_rule_only_base_model_from_the_response_model( assert cost == pytest.approx(usage.prompt_tokens * 0.0000111 + usage.completion_tokens * 0.0000222) +def _token_rates_cost(row: Mapping[str, object], usage: Usage) -> float: + input_rate: Final = row["input_cost_per_token"] + output_rate: Final = row["output_cost_per_token"] + assert isinstance(input_rate, float) and isinstance(output_rate, float), row + return usage.prompt_tokens * input_rate + usage.completion_tokens * output_rate + + +@pytest.mark.parametrize( + "requested_model", + ["accounts/fireworks/models/sol-display-test", "fireworks_ai/accounts/fireworks/models/sol-display-test"], +) +def test_completion_cost_bills_requested_model_when_response_names_an_unmapped_display_name( + _local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch, requested_model: str +) -> None: + monkeypatch.setitem( + litellm.model_cost, + "fireworks_ai/accounts/fireworks/models/sol-display-test", + {"litellm_provider": "fireworks_ai", "mode": "chat", "input_cost_per_token": 3e-6, "output_cost_per_token": 7e-6}, + ) + usage: Final = Usage(prompt_tokens=1000, completion_tokens=500, total_tokens=1500) + + cost: Final = completion_cost( + completion_response=ModelResponse(model="Sol Display Max", choices=[], usage=usage), + model=requested_model, + custom_llm_provider="fireworks_ai", + ) + + assert cost == pytest.approx( + _token_rates_cost(litellm.model_cost["fireworks_ai/accounts/fireworks/models/sol-display-test"], usage) + ) + + +def test_completion_cost_bills_requested_model_when_response_row_has_no_rates( + _local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setitem( + litellm.model_cost, + "openai/sol-requested-test", + {"litellm_provider": "openai", "mode": "chat", "input_cost_per_token": 3e-6, "output_cost_per_token": 7e-6}, + ) + monkeypatch.setitem( + litellm.model_cost, + "openai/sol-served-snapshot-test", + {"litellm_provider": "openai", "mode": "chat", "input_cost_per_token": None, "output_cost_per_token": None}, + ) + usage: Final = Usage(prompt_tokens=1000, completion_tokens=500, total_tokens=1500) + + cost: Final = completion_cost( + completion_response=ModelResponse(model="sol-served-snapshot-test", choices=[], usage=usage), + model="openai/sol-requested-test", + custom_llm_provider="openai", + ) + + assert cost == pytest.approx(_token_rates_cost(litellm.model_cost["openai/sol-requested-test"], usage)) + + +def test_completion_cost_bills_priced_response_model_over_priced_requested_model( + _local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setitem( + litellm.model_cost, + "openai/sol-requested-test", + {"litellm_provider": "openai", "mode": "chat", "input_cost_per_token": 3e-6, "output_cost_per_token": 7e-6}, + ) + monkeypatch.setitem( + litellm.model_cost, + "openai/sol-requested-test-2099-01-01", + {"litellm_provider": "openai", "mode": "chat", "input_cost_per_token": 5e-6, "output_cost_per_token": 11e-6}, + ) + usage: Final = Usage(prompt_tokens=1000, completion_tokens=500, total_tokens=1500) + + cost: Final = completion_cost( + completion_response=ModelResponse(model="sol-requested-test-2099-01-01", choices=[], usage=usage), + model="openai/sol-requested-test", + custom_llm_provider="openai", + ) + + assert cost == pytest.approx(_token_rates_cost(litellm.model_cost["openai/sol-requested-test-2099-01-01"], usage)) + + +def test_completion_cost_keeps_response_pricing_when_requested_alias_has_no_rates( + _local_model_cost_map: None, monkeypatch: pytest.MonkeyPatch +) -> None: + size_bucket: Final = get_base_model_for_pricing("Sol 70B Display") + monkeypatch.setitem( + litellm.model_cost, + size_bucket, + {"litellm_provider": "fireworks_ai", "mode": "chat", "input_cost_per_token": 3e-6, "output_cost_per_token": 7e-6}, + ) + usage: Final = Usage(prompt_tokens=1000, completion_tokens=500, total_tokens=1500) + + cost: Final = completion_cost( + completion_response=ModelResponse(model="Sol 70B Display", choices=[], usage=usage), + model="sol-team-alias", + custom_llm_provider="fireworks_ai", + ) + + assert cost == pytest.approx(_token_rates_cost(litellm.model_cost[size_bucket], usage)) + + def test_completion_cost_orders_candidates_without_extra_huggingface_config_requests( _local_model_cost_map: None, ) -> None: