Russian text in neural networks is often more expensive than English because models charge by token, and the tokenizers in many models break Russian words into more pieces. The difference between models is substantial: in a Habr measurement published on October 1, 2026, the same Russian text used 3.9 times more tokens in Claude Opus 5 than in GPT-5.5. At the same per-token price, the bill would be nearly four times higher. That is why models for Russian-language tasks should not be compared by price per million tokens, but by cost per 1,000 characters or per a typical prompt on your own data.
This material was prepared by AIrassvet based on publicly available sources, verified on October 8, 2026. It is intended for owners and finance leaders, CIOs, and product teams budgeting for AI or choosing a model for a chatbot, support, document workflows, and AI agents. Prices in the article are listed according to sources available on the publication dates and change quickly; AIrassvet is not affiliated with the model developers mentioned, and the section at the end describes the publisher's services.
Contents
- What a token is and why you pay for it
- What the 11-tokenizer benchmark showed
- In short: how much “heavier” Russian is than English
- Price per 1,000 characters instead of price per token
- Why this matters for a company budget
- How to calculate costs on your own data
- Seven ways to reduce the cost of Russian text
- What not to do
- Frequently asked questions
- Original sources and verification method
What a token is and why you pay for it
A token is a chunk of text that a model's tokenizer breaks your prompt and response into before processing. It can be a whole word, part of a word, or even a single letter. Model developers publish prices per million input and output tokens, and the bill is calculated in those terms.
The problem is that the token is different for every model. Tokenizer vocabularies were mostly trained on English text, so most models split English in roughly the same way. Russian is different: a model whose vocabulary contains little Cyrillic breaks words into several chunks or almost letter by letter. The same meaning turns into a larger number of tokens — and a larger bill.
This is not just about money. The context window is also measured in tokens, so the stated limit fits less Russian text than English: a long contract or knowledge base in Russian will hold fewer pages than the spec suggests.
What the 11-tokenizer benchmark showed
On October 1, 2026, Habr published a benchmark of 11 tokenizers, behind which are GPT, Claude, Gemini, DeepSeek, Qwen, GLM, Kimi, Grok, GigaChat, and YandexGPT. The author ran four pairs of parallel texts through them — the Russian original and an English translation of the same meaning. For open models, tokens were counted with a local tokenizer; for closed models, they were measured through differences in counter values in API responses. On October 5, CNews summarized the results. CNews.
Key takeaways from the benchmark:
- English is almost the same across all models: about 5 characters per token for 10 of the 11 models.
- Russian varies fivefold: from about 1 character per token in Claude Opus 5 to about 5 in YandexGPT and GigaChat.
- Models built for Russian have the advantage: in GigaChat and YandexGPT, Russian text actually takes fewer tokens than its English translation.
- The shift from GPT-4 to GPT-4o nearly halved Russian token usage: from 2.29 to 4.01 characters per token. If you budgeted based on GPT-4 experience, you were overestimating.
- The difference is smaller in code: Latin characters and keywords tokenize normally; mainly Russian comments become more expensive.
The author states the benchmark's limitations: there were only a few texts (about 2,500 characters of prose per language), some closed models were measured through intermediaries, and ruble prices were medians across services on a single date. Exact coefficients on your own data will differ, but according to the author the model rankings are stable across genres.
In short: how much “heavier” Russian is than English
| Models | How much longer Russian text is than English text with the same meaning (in tokens) |
|---|---|
| YandexGPT, GigaChat | Russian is shorter than English |
| GPT-5.x/6, Gemini, Grok, GLM | 18–23% longer |
| DeepSeek | 38% longer |
| Qwen | 52% longer |
| Kimi | 86% longer |
| Claude Opus 5 | Almost three times longer |
Source: Habr, October 1, 2026, three prose genres without code.
Practical takeaway: budget estimates based on English for a Russian-language product need to be multiplied by a language factor. The benchmark author suggests the following rules of thumb: about 20% more for GPT, Gemini, and Grok; 40–50% more for DeepSeek and Qwen; and about three times more for Claude Opus 5.
Price per 1,000 characters instead of price per token
To compare models fairly, the benchmark author converted prices to rubles per 1,000 characters of Russian text. The price per million tokens is the median input price across services accessible from Russia as of September 27, 2026.
- Claude Opus 5 and GPT-5.5 have the same list price, but for Russian text Opus 5 is 3.9 times more expensive: 0.53 RUB versus 0.14 RUB per 1,000 characters of input text. For English text, the difference is only 1.55 times.
- GigaChat 2 Max is 2.3 times more expensive than GPT-5.5 per token, but only 1.9 times more expensive per 1,000 Russian characters.
- YandexGPT 5 Lite is half the price of GPT-5.5 per token, but 2.6 times cheaper per 1,000 Russian characters.
There is an important caveat: the same model can have prices that differ by multiples across providers. According to the same benchmark, on September 27, 1 million input tokens for Opus 5 cost 28.5 RUB with the cheapest provider and 549 RUB at the median. So the ruble figures in the table are a tool for comparing models against each other, and for your own budget you should plug in your provider's price. The language multiplier does not change: the tokenizer built into the model, not the access reseller, determines how many tokens your Russian text will use.
Another detail: the tokenizer is the same for the response as well. Output tokens are usually 3–5 times more expensive than input tokens, so for a chatbot that writes a lot in Russian, the effect is stronger than in a calculation based on input only.
The tokenizer can also change within a single model family. On October 7, 2026, Anthropic introduced Claude Haiku 5.5cut its price to $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens, and also noted that the model has an updated tokenizer that makes some tasks use slightly more tokens. Token price cuts and token growth should be evaluated togetherand it’s worth checking this on Russian text.
Why this matters for a company’s budget
October 8, 2026 CNews published an AIrassvet scenario estimate: the 2,000 largest Russian companies could consume about 912.5 trillion tokens in 2026, which is roughly 237 billion rubles. There is no public corporate usage statistics, so this is a modeled estimate with assumptions: on average 500,000 LLM requests per day per company, 2,500 tokens per request, and an average price of 260 rubles per million tokens.
The same estimate also shows where costs are rising:
- a standard request in a corporate chat — about 2,500 tokens, an agent task with six calls to the model — about 34,000, roughly 14 times more;
- according to AIrassvet calculations based on September 2026 rates, a typical customer reply costs about 0.04 rubles, knowledge base search — 0.19 rubles, document processing — 0.73 rubles, and an AI agent task — 1.78 rubles;
- in a 2027 scenario, consumption could rise by 25–50% if 2–4% of routine requests are replaced by agent tasks.
Agents repeatedly send context back into the model — instructions, history, documents. If that context is in Russian, the language multiplier applies at every agent step. A company that chose a model with a poor tokenizer for Russian pays for it on every request.
How to calculate costs using your own data
According to the study author, measuring on your own texts takes about half an hour. The steps are:
- Collect a sample: 20–50 real requests from your scenario — the system prompt, typical user questions, excerpts from knowledge base documents, and typical answers.
- Count tokens for each candidate model. For open models, use a local tokenizer; for Claude, use Anthropic’s official free token-counting endpoint; for other closed models, infer from the difference in usage counters between two requests: with the text and without it.
- Convert to dollars: (input tokens × input price + output tokens × output price) using your provider’s rate.
- Multiply by volume: the number of requests per month, and for agents — the number of model calls per task.
- Compare with quality: a cheap model is useful only if its answers pass your checks.
| Metric | Formula |
|---|---|
| Typical request cost | input tokens × input price + output tokens × output price |
| Price per 1,000 characters | price per 1 million tokens ÷ 1,000,000 × (1,000 ÷ characters per token) |
| Monthly budget | request cost × requests per day × working days |
| Agent budget | call cost × calls per task × tasks per month |
Formulas are the AIrassvet editorial framework; all input values come from your measurement and the provider’s pricing.
Check the intermediary’s counters. In the same measurement, one aggregator’s token counters for two Claude models matched OpenAI’s tokenizer to the unit. The author suggests two explanations: the gateway is recalculating tokens itself, or requests are going to the wrong model. If your billing is based on these numbers, at least once compare them with an open tokenizer or the model’s developer dashboard.
Seven ways to lower the cost of Russian text
- Choose a model based on the cost of a typical requestrather than the list price per million tokens. A measurement can change the model choice.
- Write the service layer in Englishif the model tokenizes Russian poorly: the system prompt, tool descriptions, JSON schemas. The user text and the response remain in Russian. The measurement author notes that for Opus 5 this cuts service-layer input tokens by nearly three times, but the quality of Russian responses after such a substitution still needs to be checked.
- Route requests: high-volume and simple ones to a cheap model with strong Russian tokenization, complex ones to a stronger model. For more detail, see the article on AI model routing.
- Cache repeating context. Reading cached tokens again from many providers costs several times less; for example, with Claude Haiku 5.5, reading from cache costs $0.01 per million tokens versus $0.10 for a standard input.
- Reduce agent context: pass only the needed document excerpts through knowledge base search, not the entire document at every step. More on this in the article about controlling AI costs as token usage grows.
- Limit response length where a long answer is not needed: classification, field extraction, short customer replies.
- Calculate your own stack for large, stable volumes: an open model on your own servers shifts the per-token fee to hardware and maintenance costs. See the comparison in cloud vs. local LLM: total cost of ownership.
What not to do
- Do not translate user data into English to save money without checking quality: translation adds a model call, errors, and latency, and the savings can disappear.
- Do not compare subscriptions and API access directly. SemiAnalysis analysts, in a comparison of subscriptions from different providers, emphasize that the “value” of the same plan depends on the model and workload type, and providers can change subscription limits at any time (SemiAnalysis).
- Do not budget for a year based on a single price. Prices and tokenizers change with every new model; the measurement should be repeated whenever you switch versions.
Frequently asked questions
Why does Russian text cost more in neural networks than English?
Neural networks charge for tokens, and the tokenizers of many models were trained mainly on English text and split Russian words into more pieces. The same meaning in Russian turns into more tokens, so with the same price per token the bill is higher.
How much more expensive is Russian text in Claude than in GPT?
According to a Habr measurement from October 1, 2026, Russian text in Claude Opus 5 uses 3.9 times more tokens than in GPT-5.5. At the same token price, processing the same text on Opus 5 costs almost four times more; for English text, the difference is about 1.55 times.
Which models tokenize Russian most efficiently?
In the same measurement, YandexGPT and GigaChat tokenized Russian best: in their case, Russian text uses fewer tokens than the English translation. For GPT, Gemini, Grok, and GLM, Russian is 18–23% longer than English; for DeepSeek, 38%; for Qwen, 52%; and for Kimi, 86%.
How Do You Estimate How Many Tokens Russian Text Will Use?
Take 20–50 real prompts from your use case and count tokens for each model: use a local tokenizer for open models, the official token-count endpoint for Claude, or compare usage counters across two requests for other closed models. Then convert that into dollars using your provider’s pricing.
How Can You Reduce LLM Costs for a Russian-Language Product?
Choose the model based on the cost of a typical request on your own data, write the boilerplate part of the prompt in English if the model tokenizes Russian poorly, route simple requests to a low-cost model, cache repeating context, and shorten agent context.
Does Tokenization Affect a Model’s Context Window?
Yes. The context window is measured in tokens, so for a model that splits Russian text into more tokens, noticeably less Russian text fits into the stated limit than English. For long documents and knowledge bases in Russian, this needs to be considered when choosing a model.
Source Materials and Verification Method
Verified on October 8, 2026:
- Habr, October 1, 2026: “Russian text in Claude Opus 5 costs 3.9 times more than in GPT-5.5” — measurement of 11 tokenizers, conversion to rubles per 1,000 characters, methodology, and limitations.
- CNews, October 5, 2026 — summary of the measurement.
- Anthropic, Claude Haiku 5.5 — Haiku 5.5 pricing and a note about the updated tokenizer.
- CNews, October 8, 2026 — a scenario-based estimate by Orion soft of token consumption by large businesses in 2026–2027.
- SemiAnalysis — a comparison of subscriptions and how their value depends on the model and workload.
All tokenization coefficients and ruble prices are taken from the Habr author’s measurement on their own texts and as of the publication date; AIrassvet did not conduct its own measurement. The Orion soft estimate is a scenario model with assumptions, not statistics. The calculation approach, formulas, and savings methods are AIrassvet’s editorial analysis.
How AIrassvet Helps Reduce LLM Costs for Russian
If your model bill is growing faster than the value you get, or you are choosing a model for a new workflow, the task can be broken down step by step. On the AIrassvet website, which also uses the name “AI Studio by Vladimir Lomtev,” the following relevant services are described:
- AI use-case audit: measuring the token count and cost of a typical request on your Russian-language data across several models.
- Integrating Russian and foreign models through a unified layer that routes requests by complexity and cost.
- RAG system for a corporate knowledge base so the model receives only the needed document fragments, not the full text.
- Deployment and fine-tuning of an open model on company servers when request volume is large and stable.
The first step is to choose one process, describe its current metrics, data sources, constraints, and the criterion you will use to compare models against each other.
Bottom line: calculate rubles per request, not price per token
A price per million tokens says nothing about the cost of Russian text: different models break it into tokens differently, and the gap can reach four times. As agentic tasks grow, where context is sent back to the model repeatedly, that multiplier becomes a meaningful expense line.
The solution is simple and inexpensive: spend half an hour measuring on your own data, calculate the cost of a typical request, and route requests between models. It is worth doing before choosing a model for a new workflow and repeating it every time you switch versions.