Write Like It's 1866: LLMs Relearn Telegraphese

When instructed to communicate in "cablese"—the highly compressed dialect used by historical telegraph operators—modern Large Language Models (LLMs) can reduce their billed output tokens by 40–49%. Remarkably, data across four different model families shows that the original information is recovered with full fidelity.

To quantify this, the Telegraph Test benchmark was developed to measure both the efficiency of this compression and the subsequent ability of models to retrieve the original data.

---

Publication Details:

• Date: October 4, 2026
• Reading Time: $\approx 12$ minutes
• Word Count: $\approx 2,386$ words
• Author: Travis Smith
• Reproducibility: All tools, frozen ledgers, and a one-click notebook are available at github.com/Travis42/telegraph-test.

---

The Era of Metered Words

In 1866, the debut of the transatlantic cable introduced a staggering cost for communication: **$100 for a ten-word message**. This broke down to$10 per word (with a ten-word minimum), which equates to roughly $2,600 in contemporary currency.

Because telegraph companies billed per word, a specialized industry emerged to minimize costs.

![Image: An oil painting from the National Maritime Museum showing the Great Eastern laying the first successful Atlantic cable. Public Domain.] The Great Eastern: The vessel that established the $10-per-word link, forcing a generation to master cablese.

Two Paths to Compression

The pressure of the "per-word" price tag led to two distinct strategies:

1. Codebooks: Commercial dictionaries (such as Bentley’s ABC Telegraphic Code) spanned thousands of pages. A single word like OTTER could represent a complex phrase: "steamship arrived, cargo intact, remit balance." This allowed a 40-word business negotiation to be condensed into just 6 words.
2. Cablese: This was a disciplined, emergent writing style. A typical message looked like: "ARRIVE TUESDAY BRING FUNDS STOP CONFIRM WIFE SAILS FRIDAY." It stripped the language to its bare essentials without losing the core meaning.

While these methods vanished with the advent of the telephone, fax, and email, the spirit of telegraphese persists wherever bandwidth or cost is restricted—most notably in the 160-character limits of early SMS and the original constraints of Twitter.

---

Modern Tokens: The New Telegraph Bill

In the current AI landscape, an LLM API bill is essentially a modern telegraph bill. To test if LLMs could replicate historical savings, two strategies were applied:

• The Codebook Approach: Substituting text with words from a fixed dictionary. This proved inefficient because most prose isn't "dictionary-shaped," and it cannot remove existing words—only rename them.
• The Cablese Approach: Using a specific system prompt:

  Write a complete record in telegraphese; drop articles and filler; abbreviate; keep every fact, number, and proper noun verbatim — in lowercase, not all caps.

The Results: Token Savings and Recovery

The benchmark utilized 50 passages and approximately 1,300 questions. The following matrix shows "readers" (models extracting info from records) and "writers" (models creating the records).

Table 1: Cross-Family Performance Matrix

Note: Recovery ratios are relative to the plaintext control (1.00). All GLM calls were made via the z.ai coding plan.

Key Observations:

• The "Lowercase" Hack: If the "lowercase" instruction is omitted, models default to ALL CAPS. This styling is expensive, costing an additional 14–19 points (e.g., qwen drops to 29.8% savings).
• The GPT Anomaly: gpt-5-mini cannot disable its internal reasoning process; the act of compressing actually makes it "think" more, resulting in bills that are nearly double the plaintext cost.
• Fidelity: No comparison favored plaintext. Recovery ratios stayed between $0.99$ and $1.10$.

---

The Condition Ladder

To isolate variables, the researchers used a "condition ladder" to see how different formats affected accuracy.

The Testing Sequence:

• Plaintext Passage $\rightarrow$ Answer in plaintext $\rightarrow$ 91.5% Acc (Reference)
• Plaintext Passage $\rightarrow$ Answer in cablese $\rightarrow$ 74.4% Acc (Ratio: 0.81)
  • $\rightarrow$ Transcribe terse answer to plaintext $\rightarrow$ 78.8% Acc (Ratio: 0.86)
• Plaintext Record $\rightarrow$ Plaintext summary $\rightarrow$ 75.8% Acc (Reference)
• Cablese Record $\rightarrow$ Compressed summary $\rightarrow$ 82.5% Acc (Ratio: 1.09)
• Decoded Record $\rightarrow$ Cablese expanded back $\rightarrow$ 81.6% Acc (Ratio: 1.08)

Analysis of Findings

"The register is not a construct we invented... the capability was already in the weights, inherited from a century and a half of people writing under metered bandwidth."

1. Inherent Knowledge: Downstream models—even those that had never seen an example of cablese—could read compressed records as well as or better than plaintext.
2. The "Noise" Benefit: Cablese records actually performed slightly better (1.09 ratio) than regular responses. This is likely because compression removes the "copy-the-record" fluff that strict grading systems often penalize.
3. The Cost of Human Readability: While expanding cablese back into plaintext for humans costs some of the savings, it still remains significantly cheaper than storing the original plaintext.

Final Verdict

The data is statistically significant (in-family read-back $p < 0.0001$; decoded-record $p = 0.03$). By adopting a telegraphese register, users can achieve:

• Full price API bills $\rightarrow$ 40–49% reduction in costs.
• Zero to positive gain in information recovery.
• Efficient archiving that remains cheaper than plaintext even after expansion.