Understanding Tone-Dependent Inference Cost in Large Language Models
2026-07-27 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors studied how the tone of prompts given to large language models (LLMs) affects both the correctness of their answers and the amount of computational resources they use, measured by token output length. They tested seven tones, ranging from very polite to threatening, on a 570-question dataset and found that the length of responses changed much more than accuracy did depending on tone. Some models produced shorter or longer answers based on tone, with rude tones often leading to more efficient or balanced results. Overall, prompt tone impacts not just how good the answers are but also how much effort the LLM uses to generate them.
large language modelsprompt toneinference costoutput tokensMMLU datasetaccuracyPareto-optimal frontierChatGPTGemini modelstoken consumption
Authors
Akhil Kumar, Om Dobariya
Abstract
We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output-token consumption. Experiments were performed to understand the trade-offs between accuracy and inference cost on a 570 Question MMLU dataset for LLM models prompted in seven different tones from sycophantic to threatening. Our results show that the output-token-length variation substantially exceeded accuracy variation across all models. Output-token consumption varied by up to 44.3% across tone conditions. We also analyzed the tradeoff between the accuracy of the answers and the average output token length in the reasoning process. For the ChatGPT models 4o and 5-nano, the rude tone is quite dominant. For the Gemini models 2.5 Flash and 2.5 Flash Lite, the rude and neutral tones are dominant on the Pareto-optimal frontier. We find that prompt tone influences not only answer quality but also the amount of billable inference resources consumed by modern LLMs.