diff --git a/litellm/llms/azure_ai/chat/transformation.py b/litellm/llms/azure_ai/chat/transformation.py index f2d405e9a17..039c462b38a 100644 --- a/litellm/llms/azure_ai/chat/transformation.py +++ b/litellm/llms/azure_ai/chat/transformation.py @@ -17,6 +17,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( from litellm.llms.azure.common_utils import BaseAzureLLM from litellm.llms.azure_ai.common_utils import is_foundry_model_inference_base from litellm.llms.base_llm.chat.transformation import LiteLLMLoggingObj +from litellm.llms.openai.chat.gpt_5_transformation import OpenAIGPT5Config from litellm.llms.openai.common_utils import drop_params_from_unprocessable_entity_error from litellm.llms.openai.openai import OpenAIConfig from litellm.llms.xai.chat.transformation import XAIChatConfig @@ -42,12 +43,37 @@ NON_OPENAI_SPEC_MESSAGE_FIELDS: Final = ( ) +class AzureAIGPT5Config(OpenAIGPT5Config): + @classmethod + def _model_map_lookup_name(cls, model: str) -> str: + """Normalise a Foundry routing name to its cost-map key, when the map has one. + + A Foundry deployment and its OpenAI-hosted namesake are different products with + different capabilities, so ``azure_ai/`` is the entry to read whenever the map + carries it. Most gpt-5-family names have no ``azure_ai/`` row, though, and prefixing + those anyway costs them every flag: ``get_llm_provider`` re-resolves an ``azure_ai/`` + name to the azure provider when a global AZURE_AI_API_BASE points at an + openai.azure.com host, ``azure/`` is not a key either, so the lookup lands + nowhere and every effort answer degrades to False. A missing key defers to the base + resolver instead. + """ + prefixed: Final = model if model.startswith("azure_ai/") else f"azure_ai/{model}" + return prefixed if prefixed in litellm.model_cost else super()._model_map_lookup_name(model) + + +azureAIGPT5Config: Final = AzureAIGPT5Config() + + class AzureAIStudioConfig(OpenAIConfig): def get_supported_openai_params(self, model: str) -> list: model_supports_tool_choice = True # azure ai supports this by default if not supports_tool_choice(model=f"azure_ai/{model}"): model_supports_tool_choice = False - supported_params = super().get_supported_openai_params(model) + supported_params = ( + azureAIGPT5Config.get_supported_openai_params(model) + if azureAIGPT5Config.is_model_gpt_5_model(model) + else super().get_supported_openai_params(model) + ) if not model_supports_tool_choice: filtered_supported_params: Final = [] for param in supported_params: @@ -61,6 +87,27 @@ class AzureAIStudioConfig(OpenAIConfig): return supported_params + def map_openai_params( + self, + non_default_params: dict[str, object], # mutable-ok: OpenAIConfig.map_openai_params signature + optional_params: dict[str, object], # mutable-ok: OpenAIConfig.map_openai_params signature + model: str, + drop_params: bool, + ) -> dict[str, object]: # mutable-ok: OpenAIConfig.map_openai_params signature + if not azureAIGPT5Config.is_model_gpt_5_model(model): + return super().map_openai_params( + non_default_params=non_default_params, + optional_params=optional_params, + model=model, + drop_params=drop_params, + ) + return azureAIGPT5Config.map_openai_params( + non_default_params=non_default_params, + optional_params=optional_params, + model=model, + drop_params=drop_params, + ) + def _supports_stop_reason(self, model: str) -> bool: """ Check if the model supports stop tokens. diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 2c55828c1e2..6c6e65927f9 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -3485,6 +3485,55 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "azure_ai/gpt-6-astra": { + "cache_creation_input_token_cost": 1.25e-05, + "cache_creation_input_token_cost_above_272k_tokens": 2.5e-05, + "cache_read_input_token_cost": 1e-06, + "cache_read_input_token_cost_above_272k_tokens": 2e-06, + "input_cost_per_token": 1e-05, + "input_cost_per_token_above_272k_tokens": 2e-05, + "litellm_provider": "azure_ai", + "max_input_tokens": 922000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 5e-05, + "output_cost_per_token_above_272k_tokens": 7.5e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.01, + "search_context_size_low": 0.01, + "search_context_size_medium": 0.01 + }, + "source": "https://ai.azure.com/catalog/models/gpt-6-astra", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_computer_use": true, + "supports_function_calling": true, + "supports_max_reasoning_effort": false, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_cache_breakpoint": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, "azure_ai/gpt-5.5": { "deprecation_date": "2027-10-26", "cache_read_input_token_cost": 5e-07, @@ -7189,7 +7238,7 @@ ], "supports_computer_use": true, "supports_function_calling": true, - "supports_max_reasoning_effort": true, + "supports_max_reasoning_effort": false, "supports_minimal_reasoning_effort": false, "supports_native_streaming": true, "supports_none_reasoning_effort": true, @@ -7455,7 +7504,7 @@ ], "supports_computer_use": true, "supports_function_calling": true, - "supports_max_reasoning_effort": true, + "supports_max_reasoning_effort": false, "supports_minimal_reasoning_effort": false, "supports_native_streaming": true, "supports_none_reasoning_effort": true, diff --git a/litellm/router_utils/reasoning_effort_capability.py b/litellm/router_utils/reasoning_effort_capability.py index 9185d901a28..7b145c15a07 100644 --- a/litellm/router_utils/reasoning_effort_capability.py +++ b/litellm/router_utils/reasoning_effort_capability.py @@ -10,8 +10,8 @@ opt-in. none is opt-out everywhere except the azure gpt-5 family, whose config r UnsupportedParamsError without an explicit true. xhigh is gated on the request path by the openai and azure gpt-5 configs. max is not gated there at -all: every entry carrying supports_max_reasoning_effort is Claude-family, and -anthropic/chat/transformation.py gates max on the output_config path while its reasoning_effort +all: outside the gpt-6-astra rows every entry carrying supports_max_reasoning_effort is Claude-family, +and anthropic/chat/transformation.py gates max on the output_config path while its reasoning_effort path maps any level to a thinking budget. Making max opt-in is a deliberate trade, then, since an explicit flag is the only signal that the tier is a real one rather than litellm rounding the level to a budget, and a missing flag costs advisory metadata rather than a rejected request. diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 2c55828c1e2..6c6e65927f9 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -3485,6 +3485,55 @@ "supports_response_schema": true, "supports_tool_choice": true }, + "azure_ai/gpt-6-astra": { + "cache_creation_input_token_cost": 1.25e-05, + "cache_creation_input_token_cost_above_272k_tokens": 2.5e-05, + "cache_read_input_token_cost": 1e-06, + "cache_read_input_token_cost_above_272k_tokens": 2e-06, + "input_cost_per_token": 1e-05, + "input_cost_per_token_above_272k_tokens": 2e-05, + "litellm_provider": "azure_ai", + "max_input_tokens": 922000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 5e-05, + "output_cost_per_token_above_272k_tokens": 7.5e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.01, + "search_context_size_low": 0.01, + "search_context_size_medium": 0.01 + }, + "source": "https://ai.azure.com/catalog/models/gpt-6-astra", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_computer_use": true, + "supports_function_calling": true, + "supports_max_reasoning_effort": false, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_cache_breakpoint": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, "azure_ai/gpt-5.5": { "deprecation_date": "2027-10-26", "cache_read_input_token_cost": 5e-07, @@ -7189,7 +7238,7 @@ ], "supports_computer_use": true, "supports_function_calling": true, - "supports_max_reasoning_effort": true, + "supports_max_reasoning_effort": false, "supports_minimal_reasoning_effort": false, "supports_native_streaming": true, "supports_none_reasoning_effort": true, @@ -7455,7 +7504,7 @@ ], "supports_computer_use": true, "supports_function_calling": true, - "supports_max_reasoning_effort": true, + "supports_max_reasoning_effort": false, "supports_minimal_reasoning_effort": false, "supports_native_streaming": true, "supports_none_reasoning_effort": true, 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 df680b7cb0e..40abb5bfca3 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 @@ -2008,7 +2008,14 @@ def test_generic_cost_per_token_azure_gpt56(_local_model_cost_map, assert round(completion_cost, 10) == round(output_cost * completion_tokens, 10) -@pytest.mark.parametrize("model,zone_multiplier", [("azure/gpt-6-astra", 1.0), ("azure/us/gpt-6-astra", 1.1)]) +@pytest.mark.parametrize( + "model,custom_llm_provider,zone_multiplier", + [ + ("azure/gpt-6-astra", "azure", 1.0), + ("azure/us/gpt-6-astra", "azure", 1.1), + ("azure_ai/gpt-6-astra", "azure_ai", 1.0), + ], +) @pytest.mark.parametrize( "prompt_tokens,input_side_multiplier,output_multiplier", [(100000, 1.0, 1.0), (300000, 2.0, 1.5)], @@ -2016,6 +2023,7 @@ def test_generic_cost_per_token_azure_gpt56(_local_model_cost_map, def test_generic_cost_per_token_azure_gpt_6_astra_foundry_price_sheet( _local_model_cost_map, model, + custom_llm_provider, zone_multiplier, prompt_tokens, input_side_multiplier, @@ -2023,7 +2031,8 @@ def test_generic_cost_per_token_azure_gpt_6_astra_foundry_price_sheet( ): """Microsoft Foundry sells gpt-6-astra at the OpenAI rates: $10 input, $1 cache read, $12.50 cache write, $50 output per 1M tokens on Standard Global, with the input side doubling and output 1.5x above 272K - prompt tokens. Standard US Data Zone carries the usual 10% uplift on every rate. + prompt tokens. Standard US Data Zone carries the usual 10% uplift on every rate. A Foundry + deployment reached through the azure_ai route bills the same Standard Global sheet. """ cached_tokens = 50000 cache_write_tokens = 40000 @@ -2041,7 +2050,7 @@ def test_generic_cost_per_token_azure_gpt_6_astra_foundry_price_sheet( prompt_cost, completion_cost = generic_cost_per_token( model=model, usage=usage, - custom_llm_provider="azure", + custom_llm_provider=custom_llm_provider, ) input_side = zone_multiplier * input_side_multiplier @@ -2051,6 +2060,18 @@ def test_generic_cost_per_token_azure_gpt_6_astra_foundry_price_sheet( assert completion_cost == pytest.approx(zone_multiplier * output_multiplier * completion_tokens * 5e-5) +def test_generic_cost_per_token_azure_ai_gpt_6_astra_flex_bills_the_standard_rate(_local_model_cost_map): + usage = Usage(prompt_tokens=1000, completion_tokens=100, total_tokens=1100) + + standard = generic_cost_per_token(model="azure_ai/gpt-6-astra", usage=usage, custom_llm_provider="azure_ai") + flex = generic_cost_per_token( + model="azure_ai/gpt-6-astra", usage=usage, custom_llm_provider="azure_ai", service_tier="flex" + ) + + assert flex == standard + assert standard == pytest.approx((1000 * 1e-05, 100 * 5e-05)) + + @pytest.mark.parametrize( "model,expected_none,expected_xhigh,expected_minimal", [ diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_reasoning_effort_normalization.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_reasoning_effort_normalization.py index 56b754c3476..af7befecc33 100644 --- a/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_reasoning_effort_normalization.py +++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_handler_reasoning_effort_normalization.py @@ -82,3 +82,22 @@ class TestTheNormalizedTierIsTheTierSent: self, local_model_cost_map, model, provider, effort, expected ): assert _reasoning_effort_sent(model, provider, effort) == expected + + @pytest.mark.parametrize( + "model, provider", + [ + ("gpt-6-astra", "azure_ai"), + ("azure_ai/gpt-6-astra", "azure_ai"), + ("gpt-6-astra", "azure"), + ("us/gpt-6-astra", "azure"), + ], + ) + def test_an_azure_hosted_astra_deployment_drops_to_the_tier_it_accepts( + self, local_model_cost_map, model, provider + ): + """The deployment answers ``max`` with a 400 naming ``none`` through ``xhigh``, so the rows + say so and the adapter sends the tier below instead of the rejected one.""" + assert _reasoning_effort_sent(model, provider, "max") == "xhigh" + + def test_the_openai_hosted_twin_still_sends_max(self, local_model_cost_map): + assert _reasoning_effort_sent("gpt-6-astra", "openai", "max") == "max" diff --git a/tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py b/tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py index 33fbb4e8fc7..f8cc0b5071e 100644 --- a/tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py +++ b/tests/test_litellm/llms/azure_ai/chat/test_azure_ai_transformation.py @@ -3,6 +3,8 @@ from unittest.mock import MagicMock, patch import pytest +import litellm +from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map from litellm.llms.azure_ai.azure_model_router.transformation import ( AzureModelRouterConfig, ) @@ -138,6 +140,46 @@ def test_azure_ai_validate_environment_with_azure_ad_token(): assert headers["Content-Type"] == "application/json" +@pytest.fixture +def _local_model_cost_map(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + monkeypatch.setattr(litellm, "model_cost", get_model_cost_map(url=litellm.model_cost_map_url)) + + +def test_foundry_gpt_6_astra_keeps_sampling_params_when_reasoning_effort_is_none(_local_model_cost_map): + optional_params = AzureAIStudioConfig().map_openai_params( + non_default_params={"reasoning_effort": "none", "temperature": 0.2, "top_p": 0.9}, + optional_params={}, + model="gpt-6-astra", + drop_params=False, + ) + + assert optional_params == {"reasoning_effort": "none", "temperature": 0.2, "top_p": 0.9} + + +def test_a_gpt_5_name_without_a_foundry_row_keeps_reading_its_own_entry( + monkeypatch: pytest.MonkeyPatch, _local_model_cost_map +): + """Most gpt-5-family names have no azure_ai/ row. Reading an azure_ai/ key for those finds + nothing, and an openai.azure.com base sends the name down the azure provider, which has no key + for it either, so every effort answer would silently fall back to false and take temperature, + top_p and logprobs down with it.""" + monkeypatch.setenv("AZURE_AI_API_BASE", "https://example-resource.openai.azure.com") + monkeypatch.setenv("AZURE_AI_API_KEY", "placeholder") + + optional_params = litellm.utils.get_optional_params( + model="gpt-5.1-chat-latest", + custom_llm_provider="azure_ai", + temperature=0.2, + top_p=0.9, + logprobs=True, + ) + + assert optional_params["temperature"] == 0.2 + assert optional_params["top_p"] == 0.9 + assert optional_params["logprobs"] is True + + def test_azure_ai_grok_stop_parameter_handling(): """ Test that Grok models properly handle stop parameter filtering in Azure AI Studio. diff --git a/tests/test_litellm/proxy/auth/test_model_checks.py b/tests/test_litellm/proxy/auth/test_model_checks.py index d58683fd1e5..36bfc4c5dd3 100644 --- a/tests/test_litellm/proxy/auth/test_model_checks.py +++ b/tests/test_litellm/proxy/auth/test_model_checks.py @@ -857,6 +857,24 @@ def test_add_known_models_refreshes_models_by_provider_for_wildcard_expansion(): litellm.add_known_models(model_cost_map={}) assert fake_model not in litellm.models_by_provider["vertex_ai"] + +def test_azure_ai_wildcard_lists_the_foundry_gpt_6_astra_entry(monkeypatch): + import litellm + from litellm.proxy.auth.model_checks import get_known_models_from_wildcard + + monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + foundry_key = "azure_ai/gpt-6-astra" + local_entry = litellm.get_model_cost_map(url="")[foundry_key] + registered_before = foundry_key in litellm.azure_ai_models + try: + litellm.add_known_models(model_cost_map={foundry_key: local_entry}) + assert foundry_key in get_known_models_from_wildcard("azure_ai/*") + finally: + if not registered_before: + litellm.azure_ai_models.discard(foundry_key) + litellm.add_known_models(model_cost_map={}) + + def test_get_complete_model_list_drops_no_default_models_sentinel(): from litellm.proxy.auth.model_checks import get_complete_model_list diff --git a/tests/test_litellm/router_utils/test_reasoning_effort_capability.py b/tests/test_litellm/router_utils/test_reasoning_effort_capability.py index f181370455d..ccd6766b13a 100644 --- a/tests/test_litellm/router_utils/test_reasoning_effort_capability.py +++ b/tests/test_litellm/router_utils/test_reasoning_effort_capability.py @@ -389,14 +389,24 @@ class TestGpt6AstraAdvertisesItsDocumentedLevels: "max", ) - @pytest.mark.parametrize("model", ["azure/gpt-6-astra", "azure/us/gpt-6-astra"]) - def test_a_foundry_deployment_also_advertises_none(self, local_model_cost_map, model): - """Microsoft Foundry serves the same model but its API accepts reasoning_effort none - (verified live: 200 with zero reasoning tokens, and it unlocks temperature), which - OpenAI's rejects, so an Azure deployment offers none on top of low through max.""" + @pytest.mark.parametrize( + "model,custom_llm_provider", + [ + ("azure/gpt-6-astra", "azure"), + ("azure/us/gpt-6-astra", "azure"), + ("azure_ai/gpt-6-astra", "azure_ai"), + ], + ) + def test_an_azure_hosted_deployment_advertises_none_but_not_max( + self, local_model_cost_map, model, custom_llm_provider + ): + """Microsoft hosts the same model with a different level set than OpenAI does. Verified live + on both Azure routes: none returns 200 with zero reasoning tokens and unlocks temperature, + which OpenAI's API rejects, while max returns 400 unsupported_value naming none through + xhigh as the levels it does take.""" from litellm.utils import _get_model_info_helper - model_info = dict(_get_model_info_helper(model=model, custom_llm_provider="azure")) + model_info = dict(_get_model_info_helper(model=model, custom_llm_provider=custom_llm_provider)) assert resolve_supported_reasoning_efforts(model_info, deployment_is_mapped=True) == ( "none", @@ -404,5 +414,4 @@ class TestGpt6AstraAdvertisesItsDocumentedLevels: "medium", "high", "xhigh", - "max", )