Telegraph-style writing reportedly cut billed AI output tokens by 40–49% in three tested models

Roemmele cites Travis Smith's benchmark, in which lowercase telegraph-style records used 40.4% to 48.9% fewer billed output tokens in three models. Smith reports that other models answered short questions from those records about as well as from plain-English summaries. He had not tested whether simply asking for terse writing would work as well. Asking a model to answer in the compressed style while working reduced accuracy to 0.81 of plain-English performance.