New Pennsylvania University paper finds， being rude to some LLMs leads to considerably shorter responses and higher accuracy.

i.e. the way you speak to a model change your inference bill？

Researchers ran the same 570 MMLU questions in 7 tones， from sycophantic to threatening across many models （of course to note， many of the models were old）.

The questions stayed fixed， tone prefixes were similar in length， and each model received the same step-by-step reasoning instruction.

Accuracy moved by at most 2.99% points within a model， but output-token use shifted by as much as 44.3%.

The best tone was model-specific.

For ChatGPT-4o， rude produced the highest accuracy， 89.04%， and the shortest response， 223 tokens on average.

For Gemini 2.5 Flash Lite， neutral did both： 88.25% accuracy with 1，222 tokens， while rude fell to 85.26% and used 35.2% more tokens.

The authors use selected Gemini 2.5 Flash Lite cases to suggest a mechanism： rude responses took shortcuts， while neutral responses added option checks and self-correction.

Still， tone is not merely a UX choice； it is a model-specific cost and reliability setting that production systems should standardize and benchmark.

• arxiv. org/abs/2607.23915

Title： "Understanding Tone-Dependent Inference Cost in LLMs"