From ea6c18baa503ebda6bde5ae2e812091c8f54aeb5 Mon Sep 17 00:00:00 2001 From: Vairo Di Pasquale Date: Mon, 10 Aug 2026 14:07:17 -0400 Subject: [PATCH] fix(cost): price dict-shaped image input token details at image rate (#33490) calculate_image_response_cost_from_usage read input_tokens_details with getattr(), but OpenAI images.edit responses carry it as a plain dict, so both fields came back None and all input tokens were priced at input_cost_per_token instead of input_cost_per_image_token (e.g. $5/M instead of $8/M for gpt-image-2). Read it with the dict-tolerant _get_token_detail_value helper, as the output side of the same function already does. Co-authored-by: mubashir1osmani --- .../litellm_core_utils/llm_cost_calc/utils.py | 8 ++- .../llm_cost_calc/test_llm_cost_calc_utils.py | 54 +++++++++++++++++++ 2 files changed, 60 insertions(+), 2 deletions(-) diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py index 6574b261774..05cc115c0b7 100644 --- a/litellm/litellm_core_utils/llm_cost_calc/utils.py +++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py @@ -996,9 +996,13 @@ def calculate_image_response_cost_from_usage( input_tokens_details: Final = getattr(usage, "input_tokens_details", None) prompt_tokens_details: PromptTokensDetailsWrapper | None = None if input_tokens_details is not None: + # input_tokens_details may be a dict (e.g. OpenAI image edit responses) + # or an object; read it tolerantly like the output side below, so image + # input tokens are priced at input_cost_per_image_token instead of + # silently falling back to the text rate. prompt_tokens_details = PromptTokensDetailsWrapper( - text_tokens=getattr(input_tokens_details, "text_tokens", None), - image_tokens=getattr(input_tokens_details, "image_tokens", None), + text_tokens=_get_token_detail_value(input_tokens_details, "text_tokens"), + image_tokens=_get_token_detail_value(input_tokens_details, "image_tokens"), cached_tokens=0, ) 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 0970e029ad8..5cdf74cc04b 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 @@ -2446,6 +2446,60 @@ def test_token_type_cost_breakdown_applies_regional_uplift(): assert text_input_cost + eu.cache_read_cost == pytest.approx(prompt_cost) +@pytest.mark.parametrize("details_as_dict", [True, False]) +def test_image_response_input_image_tokens_priced_at_image_rate(details_as_dict): + """ + Image input tokens must be priced at input_cost_per_image_token even when + input_tokens_details is a plain dict, as in OpenAI image edit responses. + + Regression test: dict-shaped input_tokens_details was read with getattr(), + which returns None for dicts, so image input tokens silently fell back to + the text input rate (e.g. $5/M instead of $8/M for gpt-image-2). + """ + from unittest.mock import patch + + from litellm.litellm_core_utils.llm_cost_calc.utils import ( + calculate_image_response_cost_from_usage, + ) + from litellm.types.utils import Usage + + mock_model_info = { + "input_cost_per_token": 5e-6, + "input_cost_per_image_token": 8e-6, + "output_cost_per_image_token": 3e-5, + } + + input_details = {"text_tokens": 19, "image_tokens": 512} + image_response = ImageResponse(data=[ImageObject(b64_json="x")]) + # Mirror the usage shape of a real OpenAI images.edit response: + # a Usage object carrying input_tokens/output_tokens with detail dicts. + image_response.usage = Usage( + prompt_tokens=0, + completion_tokens=0, + total_tokens=689, + input_tokens=531, + input_tokens_details=( + input_details + if details_as_dict + else ImageUsageInputTokensDetails(**input_details) + ), + output_tokens=158, + output_tokens_details={"image_tokens": 158, "text_tokens": 0}, + ) + + with patch( + "litellm.litellm_core_utils.llm_cost_calc.utils.get_model_info", + return_value=mock_model_info, + ): + cost = calculate_image_response_cost_from_usage( + model="gpt-image-2", + image_response=image_response, + custom_llm_provider="openai", + ) + + expected = 19 * 5e-6 + 512 * 8e-6 + 158 * 3e-5 + assert cost is not None + assert round(cost, 12) == round(expected, 12) GEMINI_DAY0_LAUNCH_PRICING = [ ("gemini-3.6-flash", 1.5e-06, 7.5e-06, 1.5e-07), ("gemini/gemini-3.6-flash", 1.5e-06, 7.5e-06, 1.5e-07),