diff --git a/tests/llm_responses_api_testing/base_responses_api.py b/tests/llm_responses_api_testing/base_responses_api.py index 3f77bf0600..fd39c13604 100644 --- a/tests/llm_responses_api_testing/base_responses_api.py +++ b/tests/llm_responses_api_testing/base_responses_api.py @@ -181,7 +181,7 @@ class BaseResponsesAPITest(ABC): # basic test assert the usage seems reasonable print("response_completed_event.response.usage=", response_completed_event.response.usage) assert response_completed_event.response.usage.input_tokens > 0 and response_completed_event.response.usage.input_tokens < 100 - assert response_completed_event.response.usage.output_tokens > 0 and response_completed_event.response.usage.output_tokens < 100 + assert response_completed_event.response.usage.output_tokens > 0 and response_completed_event.response.usage.output_tokens < 1000 assert response_completed_event.response.usage.total_tokens > 0 and response_completed_event.response.usage.total_tokens < 1000 # total tokens should be the sum of input and output tokens