Write Like It's 1866: LLMs Relearn Telegraphese
A benchmark finds LLMs can store records in telegraph-style cablese, cutting tokens by up to 49 percent without hurting later answers.
A reproducible Telegraph Test on about 50 passages and 1,300 questions finds models can write and read telegraph-style cablese at parity with plaintext. Writer token savings ranged from about 18 percent for gpt-5-mini to 48.9 percent for qwen3.8-27b, while GLM-5.3-Flash saved 48.4 percent with a lowercase instruction. Readers from several families, including models never shown an example, answered from compressed records at recovery ratios of 0.99 to 1.10 versus plaintext. Compressing unfinished answers hurts accuracy, but compressing settled records for later model use does not; gpt-5-mini is an exception because mandatory reasoning makes its writes cost about double.