From 2548e960f18d5260fc4fb511f2d67406cb8ce8f7 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 26 Aug 2026 14:43:21 -0700 Subject: [PATCH] fix(cost-map): correct Gemini TTS and native-audio rates Gemini 2.5 Flash Preview TTS, Gemini 2.5 Pro Preview TTS, and the three gemini-2.5-flash-native-audio entries carried rates copied from the text models, so audio output was billed 2x to 6x under Google's published prices. Set the published per-token rates on all ten keys, add output_cost_per_audio_token to the native-audio entries, and drop the long-context tier rates Google does not publish for Pro TTS. --- ...odel_prices_and_context_window_backup.json | 84 +++++----- model_prices_and_context_window.json | 84 +++++----- .../test_gemini_tts_native_audio_pricing.py | 144 ++++++++++++++++++ 3 files changed, 228 insertions(+), 84 deletions(-) create mode 100644 tests/test_litellm/test_gemini_tts_native_audio_pricing.py diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index d9a13ef1b98..467acce9850 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -21091,18 +21091,15 @@ }, "gemini-2.5-pro-preview-tts": { "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, + "input_cost_per_token": 1e-06, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview", + "output_cost_per_token": 2e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_modalities": [ "text" ], @@ -22212,11 +22209,11 @@ "supports_image_size": false }, "gemini/gemini-2.5-flash-preview-tts": { - "input_cost_per_token": 3e-07, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "mode": "audio_speech", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_token": 1e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -23156,19 +23153,16 @@ }, "gemini/gemini-2.5-pro-preview-tts": { "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, + "input_cost_per_token": 1e-06, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, + "output_cost_per_token": 2e-05, "rpm": 10000, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_modalities": [ "text" ], @@ -48804,15 +48798,16 @@ } }, "gemini-2.5-flash-native-audio-latest": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48829,15 +48824,16 @@ "gemini_native_audio": true }, "gemini-2.5-flash-native-audio-preview-09-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48854,15 +48850,16 @@ "gemini_native_audio": true }, "gemini-2.5-flash-native-audio-preview-12-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48912,15 +48909,16 @@ "gemini_audio_only_live": true }, "gemini/gemini-2.5-flash-native-audio-latest": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48939,15 +48937,16 @@ "gemini_native_audio": true }, "gemini/gemini-2.5-flash-native-audio-preview-09-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48966,15 +48965,16 @@ "gemini_native_audio": true }, "gemini/gemini-2.5-flash-native-audio-preview-12-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -49043,11 +49043,11 @@ "rpm": 10 }, "gemini-2.5-flash-preview-tts": { - "input_cost_per_token": 3e-07, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "mode": "audio_speech", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_token": 1e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index d9a13ef1b98..467acce9850 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -21091,18 +21091,15 @@ }, "gemini-2.5-pro-preview-tts": { "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, + "input_cost_per_token": 1e-06, "litellm_provider": "vertex_ai-language-models", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview", + "output_cost_per_token": 2e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_modalities": [ "text" ], @@ -22212,11 +22209,11 @@ "supports_image_size": false }, "gemini/gemini-2.5-flash-preview-tts": { - "input_cost_per_token": 3e-07, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "mode": "audio_speech", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_token": 1e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ], @@ -23156,19 +23153,16 @@ }, "gemini/gemini-2.5-pro-preview-tts": { "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, + "input_cost_per_token": 1e-06, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 65535, "max_tokens": 65535, "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, + "output_cost_per_token": 2e-05, "rpm": 10000, - "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview", + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_modalities": [ "text" ], @@ -48804,15 +48798,16 @@ } }, "gemini-2.5-flash-native-audio-latest": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48829,15 +48824,16 @@ "gemini_native_audio": true }, "gemini-2.5-flash-native-audio-preview-09-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48854,15 +48850,16 @@ "gemini_native_audio": true }, "gemini-2.5-flash-native-audio-preview-12-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48912,15 +48909,16 @@ "gemini_audio_only_live": true }, "gemini/gemini-2.5-flash-native-audio-latest": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48939,15 +48937,16 @@ "gemini_native_audio": true }, "gemini/gemini-2.5-flash-native-audio-preview-09-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -48966,15 +48965,16 @@ "gemini_native_audio": true }, "gemini/gemini-2.5-flash-native-audio-preview-12-2025": { - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, + "input_cost_per_audio_token": 3e-06, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "max_input_tokens": 1048576, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 2e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/realtime" ], @@ -49043,11 +49043,11 @@ "rpm": 10 }, "gemini-2.5-flash-preview-tts": { - "input_cost_per_token": 3e-07, + "input_cost_per_token": 5e-07, "litellm_provider": "gemini", "mode": "audio_speech", - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/pricing", + "output_cost_per_token": 1e-05, + "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/audio/speech" ] diff --git a/tests/test_litellm/test_gemini_tts_native_audio_pricing.py b/tests/test_litellm/test_gemini_tts_native_audio_pricing.py new file mode 100644 index 00000000000..88fd14e436d --- /dev/null +++ b/tests/test_litellm/test_gemini_tts_native_audio_pricing.py @@ -0,0 +1,144 @@ +import json +from pathlib import Path +from typing import Final + +import pytest + +import litellm +from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token +from litellm.types.utils import CompletionTokensDetailsWrapper, PromptTokensDetailsWrapper, Usage + +REPO_ROOT: Final = Path(__file__).parents[2] +MAIN_PATH: Final = REPO_ROOT / "model_prices_and_context_window.json" +BACKUP_PATH: Final = REPO_ROOT / "litellm" / "model_prices_and_context_window_backup.json" + +FLASH_TTS_KEYS: Final = ("gemini-2.5-flash-preview-tts", "gemini/gemini-2.5-flash-preview-tts") +PRO_TTS_KEYS: Final = ("gemini-2.5-pro-preview-tts", "gemini/gemini-2.5-pro-preview-tts") +NATIVE_AUDIO_KEYS: Final = tuple( + f"{prefix}gemini-2.5-flash-native-audio-{suffix}" + for prefix in ("", "gemini/") + for suffix in ("latest", "preview-09-2025", "preview-12-2025") +) + +FLASH_TTS_INPUT: Final = 5e-07 +FLASH_TTS_AUDIO_OUTPUT: Final = 1e-05 +PRO_TTS_INPUT: Final = 1e-06 +PRO_TTS_AUDIO_OUTPUT: Final = 2e-05 +NATIVE_AUDIO_TEXT_INPUT: Final = 5e-07 +NATIVE_AUDIO_AUDIO_INPUT: Final = 3e-06 +NATIVE_AUDIO_TEXT_OUTPUT: Final = 2e-06 +NATIVE_AUDIO_AUDIO_OUTPUT: Final = 1.2e-05 + +PUBLISHED_RATES: Final = { + **{ + key: {"input_cost_per_token": FLASH_TTS_INPUT, "output_cost_per_token": FLASH_TTS_AUDIO_OUTPUT} + for key in FLASH_TTS_KEYS + }, + **{ + key: {"input_cost_per_token": PRO_TTS_INPUT, "output_cost_per_token": PRO_TTS_AUDIO_OUTPUT} + for key in PRO_TTS_KEYS + }, + **{ + key: { + "input_cost_per_token": NATIVE_AUDIO_TEXT_INPUT, + "input_cost_per_audio_token": NATIVE_AUDIO_AUDIO_INPUT, + "output_cost_per_token": NATIVE_AUDIO_TEXT_OUTPUT, + "output_cost_per_audio_token": NATIVE_AUDIO_AUDIO_OUTPUT, + } + for key in NATIVE_AUDIO_KEYS + }, +} +ALL_KEYS: Final = tuple(PUBLISHED_RATES) +LONG_CONTEXT_TIER_FIELDS: Final = ( + "input_cost_per_token_above_200k_tokens", + "output_cost_per_token_above_200k_tokens", + "cache_read_input_token_cost_above_200k_tokens", +) + + +def _load(path: Path) -> dict: + with open(path, encoding="utf-8") as f: + return json.load(f) + + +@pytest.fixture +def local_model_cost_map(monkeypatch): + original_model_cost = litellm.model_cost + monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True") + litellm.model_cost = litellm.get_model_cost_map(url="") + litellm.get_model_info.cache_clear() + try: + yield + finally: + litellm.model_cost = original_model_cost + litellm.get_model_info.cache_clear() + + +@pytest.mark.parametrize("model", ALL_KEYS) +@pytest.mark.parametrize("path", (MAIN_PATH, BACKUP_PATH), ids=("main", "backup")) +def test_published_rates_are_registered(model: str, path: Path): + info = _load(path)[model] + for field, value in PUBLISHED_RATES[model].items(): + assert info[field] == value, f"{model} {field} in {path.name}: {info.get(field)} != {value}" + + +@pytest.mark.parametrize("model", PRO_TTS_KEYS) +@pytest.mark.parametrize("path", (MAIN_PATH, BACKUP_PATH), ids=("main", "backup")) +def test_pro_tts_has_no_long_context_tier(model: str, path: Path): + info = _load(path)[model] + for field in LONG_CONTEXT_TIER_FIELDS: + assert field not in info, f"{model} has {field} but Google publishes one flat TTS rate" + + +@pytest.mark.parametrize("model", ALL_KEYS) +def test_backup_matches_main(model: str): + assert _load(BACKUP_PATH)[model] == _load(MAIN_PATH)[model] + + +@pytest.mark.parametrize( + ("model", "provider", "input_rate", "audio_output_rate"), + ( + ("gemini-2.5-flash-preview-tts", "gemini", FLASH_TTS_INPUT, FLASH_TTS_AUDIO_OUTPUT), + ("gemini-2.5-pro-preview-tts", "gemini", PRO_TTS_INPUT, PRO_TTS_AUDIO_OUTPUT), + ("gemini-2.5-pro-preview-tts", "vertex_ai", PRO_TTS_INPUT, PRO_TTS_AUDIO_OUTPUT), + ), +) +def test_tts_audio_output_is_billed_at_the_audio_rate( + model: str, provider: str, input_rate: float, audio_output_rate: float, local_model_cost_map +): + usage: Final = Usage( + prompt_tokens=9, + completion_tokens=49, + total_tokens=58, + prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=9), + completion_tokens_details=CompletionTokensDetailsWrapper(audio_tokens=49, text_tokens=0), + ) + prompt_cost, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider=provider) + assert prompt_cost == pytest.approx(9 * input_rate) + assert completion_cost == pytest.approx(49 * audio_output_rate) + + +@pytest.mark.parametrize("model", NATIVE_AUDIO_KEYS) +def test_native_audio_output_is_billed_at_the_audio_rate(model: str, local_model_cost_map): + usage: Final = Usage( + prompt_tokens=377, + completion_tokens=84, + total_tokens=461, + prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=377), + completion_tokens_details=CompletionTokensDetailsWrapper(audio_tokens=48, reasoning_tokens=36, text_tokens=0), + ) + prompt_cost, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="gemini") + assert prompt_cost == pytest.approx(377 * NATIVE_AUDIO_TEXT_INPUT) + assert completion_cost == pytest.approx(48 * NATIVE_AUDIO_AUDIO_OUTPUT + 36 * NATIVE_AUDIO_TEXT_OUTPUT) + + +@pytest.mark.parametrize("model", NATIVE_AUDIO_KEYS) +def test_native_audio_input_is_billed_at_the_audio_rate(model: str, local_model_cost_map): + usage: Final = Usage( + prompt_tokens=1000, + completion_tokens=0, + total_tokens=1000, + prompt_tokens_details=PromptTokensDetailsWrapper(text_tokens=100, audio_tokens=900), + ) + prompt_cost, _ = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="gemini") + assert prompt_cost == pytest.approx(100 * NATIVE_AUDIO_TEXT_INPUT + 900 * NATIVE_AUDIO_AUDIO_INPUT)