# Telegraph Test reports lower token use for model-readable records

_Published Wednesday, October 7, 2026 at 9:35 AM EDT · AI · Latest · Tier 2 — Notable_

The Telegraph Test reported that language models could read compressed records written in telegraph-style English with recovery ratios of 0.99-1.10 relative to plaintext. Its benchmark used 50 passages and short factual questions, with fixed scoring rules rather than a model judge.

A lowercase instruction produced 48.4% savings for GLM-5.3-Flash on its provider's meter. The technique removes articles and filler while retaining facts, numbers and proper nouns, targeting settled records that another model will consume.

The savings did not extend to every workload. Compressing while composing answers reduced recovery, and gpt-5-mini's mandatory reasoning made compressed writes cost double despite shorter text. The author also had not tested whether generic terse instructions would produce equivalent results.

## Sources

- [fiveminutesforward.com](https://fiveminutesforward.com/post/2026-10-04-telegraph-test/)

---
Canonical: https://techandbusiness.org/newswire/wKNIeMfZo31Rv_kWwML0w8
Published: 2026-10-07T13:35:29.449Z
Story chronology: 2026-10-07T12:04:36.000Z
Retrieved: 2026-10-07T16:18:44.772Z
Publisher: Tech & Business (techandbusiness.org)