Did you take in to account that the default GPT4 model changed in June? That would explain the response difference. You can still use the original model released in march with the GPT-4-0314 identifier.