Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Typing "ghbdtn" Instead of "привет" (Hello) Costs an LLM 4–7x More: 350–500 Tokens vs 70–90

If you type "ghbdtn" instead of "привет" (hello), the neural network doesn't immediately realize it's Russian text typed on an English layout, and it spends tokens decoding it. The author of the Zyabra utility measured it: the clean phrase takes 70–90 tokens, while the mistyped one with Russian context in the chat takes 350–500, i.e. 4–7x more. Without context, the model once burned 2298 tokens and 145 seconds and still failed to decode the phrase. So it's cheaper to fix the layout before sending.

AI-processed from Habr AI; edited by Hamidun News
Typing "ghbdtn" Instead of "привет" (Hello) Costs an LLM 4–7x More: 350–500 Tokens vs 70–90
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
"Умничка)" (Good job))70 tokens and 5.4 s vs 2356 tokens and 148 s in the wrong layout
One Russian message in the history344 tokens and 23 s instead of 2298 and 145 s
Zyabra is free and localno network, no keystroke log, no clipboard access

Anyone who types a lot has sent "ghbdtn" instead of "привет" (hello) at least once: fingers on the English layout, Russian in your head. In an ordinary messenger it's a minor embarrassment. With a language model, money and time come into it: the neural network tries to work out what you meant and spends tokens on it (units of text that you pay for and that determine response speed).

The author of the Zyabra utility (a local layout switcher that erases what was typed wrongly and retypes it correctly) decided to count exactly how much. The result: a clean message costs 70–90 tokens and 5–7 seconds, while an accidental layout typo with Russian context in the chat costs 350–500 tokens and 25–30 seconds, i.e.

4–7x more.

First run: numbers that make you uneasy

The experiment is simple. The author took short phrases he could realistically send by accident in the wrong layout, and for each one sent a local model two versions: a normal one and a "broken" one. No system prompt, no chat history, a single message. The model was a local Qwen3.8-Flash-Next running on Strata, as the author calls it.

The results diverged sharply:

  • "Умничка)" (Good job)) — 70 tokens and 5.4 seconds, but in the wrong layout ("Evybxrf)") — 2356 tokens and 148 seconds, and the model never decoded the phrase;
  • "Спасибо, то что нужно!" (Thanks, just what I needed!) — 84 tokens and 7.5 seconds vs 440 tokens and 30 seconds, decoded correctly;
  • "Погнали" (Let's go) — 66 tokens and 6.2 seconds vs 336 tokens and 24 seconds, also correct.

On "Evybxrf)" the model went off trying ROT13 and the Caesar cipher, spent almost two and a half minutes, and in the end asked what was meant. It looked like a killer argument for the utility: savings of dozens of times.

Wait: in a real chat, the message is rarely the first one

The author himself noticed the weak spot in the test. In real life there is almost always a Russian-language conversation before the typo, whereas in the measurement the model solved the puzzle from scratch. He repeated the check on the same phrase in three variants:

  • no context — 2298 tokens, 145 seconds, not decoded;
  • with the system prompt "we communicate in Russian" — 495 tokens, 32 seconds, decoded, though with a minor inaccuracy ("умничк)");
  • with one previous Russian message in the history — 344 tokens, 23 seconds, decoded accurately.

Context makes the response 6–7x cheaper and faster. A real message worked even better than an explicit instruction: the model has something to latch onto, and it more quickly realizes that it's looking at Russian text typed in Latin letters rather than a cipher.

An honest bottom line and what follows from it

The realistic cost of one accidental typo is not 2000+ tokens and two and a half minutes, but roughly 350–500 tokens and 25–30 seconds when the chat already has Russian context. That is still 4–7x more than a clean message. If you make two dozen such slips in an evening, that adds up to several thousand extra tokens and a couple of minutes of waiting. Plus there is the risk that the model fails and simply asks you to clarify.

For those who pay for an API, the figures will be different, but the logic is the same: tokens go not to your question but to decoding. There are two practical takeaways. First: if you are building a chatbot or assistant for Russian-speaking users, keep Russian-language context in the prompt or history, since it noticeably lowers the cost of a typing error. Second: it is better to fix the layout before sending than to count on the model figuring it out.

There are also limitations. These are measurements by one person, on one local model, with a handful of phrases, not a broad benchmark. Other models may handle a "crooked" layout better or worse, and response time in cloud services also depends on load. The author himself calls the result "not inflated" rather than universal.

Zyabra itself is free, runs on Windows, Linux and macOS, and the code and binaries are at github.com/DmNep/zyabra. The author stresses that it is fully local: it doesn't write what you type to disk, keep a keystroke log, touch the clipboard, or open a network connection. Only the current unfinished word and the one most recent replacement are held in memory. For a program that sees every keypress, that is an important detail.

Story

Key Questions Answered

What are the key details on First run: numbers that make you uneasy?

The experiment is simple. The author took short phrases he could realistically send by accident in the wrong layout, and for each one sent a local model two versions: a normal one and a "broken" one. No system prompt, no chat history, a single message. The model was a local Qwen3.8-Flash-Next running on Strata, as the author calls it.

What are the key details on Wait: in a real chat, the message is rarely the first one?

The author himself noticed the weak spot in the test. In real life there is almost always a Russian-language conversation before the typo, whereas in the measurement the model solved the puzzle from scratch. He repeated the check on the same phrase in three variants: no context — 2298 tokens, 145 seconds, not decoded; with the system prompt "we communicate in Russian" — 495 tokens, 32 seconds, decoded, though with a minor inaccuracy ("умничк)"); with one previous Russian message in the history —…

What are the key details on An honest bottom line and what follows from it?

The realistic cost of one accidental typo is not 2000+ tokens and two and a half minutes, but roughly 350–500 tokens and 25–30 seconds when the chat already has Russian context. That is still 4–7x more than a clean message. If you make two dozen such slips in an evening, that adds up to several thousand extra tokens and a couple of minutes of waiting. Plus there is the risk that the model fails and simply asks you to clarify.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…