LLMs Relearn Telegraphese Style

A test using a 50-passage benchmark shows LLMs compress text into telegraphese style, saving 25% to 49% on costs. GLM-5.3-Flash achieves 48.4% savings with lowercase instructions. Models read compressed records at parity or better, though disabling reasoning in gpt-5-mini doubles costs. This technique relies on latent training data rather than new hardware or codebooks.