WHY THERE'S A PROBLEM WITH RUSSIAN IN THE FIRST PLACE
Russian is a language with rich morphology: cases, genders, verb aspects, free word order. An LLM trained primarily on an English corpus stumbles not on meaning, but on form. Classic symptom — the model understands the question but produces agreement at the level of "three big houses" (grammatically incorrect) or calqued phrases like "make sense" instead of "have sense".
The second layer of the problem is tokenization. Cyrillic in older-generation BPE tokenizers consumed 2-3 tokens per word, which made queries more expensive and cut context. By 2026, most frontier models switched to tokenizers where Russian costs about 1.3-1.7 tokens per word — still more expensive than English, but tolerable.
The third layer — factology and local context: Russian laws, toponyms, cultural realities, current events. Here Western models often flounder, while specialized Russian-language ones are in their element.
HOW I TESTED
Synthetic benchmarks like MMLU translate to Russian poorly and measure the wrong thing. So I ran models on five categories of practical tasks close to what developers actually do:
- Grammar and style — rewrite bureaucratic language into human text, catch agreement errors.
- Summarization — compress a 3000-word article into 5 theses without hallucinations.
- Data extraction (JSON) — pull structure from unstructured Russian text.
- Code + comments in Russian — generate a function with meaningful Russian docstrings.
- Local knowledge — questions on Russian law, geography, recent events.
Evaluation — manual, blind comparison, on a 1-10 scale. I didn't trust a single automatic judge: an LLM judge itself flounders in Russian and inflates scores for verbose answers. I track fresh model releases and wrappers in the REDDYX catalog — there you can see who shipped a new version literally this week.
RESULTS: SUMMARY TABLE
Figures are averaged subjective scores from my run, not absolute truth. Per community measurements (Ru-MMLU-like sets, MERA) the picture roughly matches.
| Model | Grammar | Summarization | JSON extraction | Local knowledge | Price (relative) | Data in RF |
|---|---|---|---|---|---|---|
| Claude (Opus/Sonnet) | 9.5 | 9.5 | 9.5 | 7.0 | High | No |
| GPT (flagship) | 9.0 | 9.0 | 9.0 | 7.5 | High | No |
| DeepSeek | 8.5 | 8.5 | 9.0 | 6.5 | Very low | No |
| YandexGPT | 8.5 | 8.0 | 7.5 | 9.5 | Medium | Yes |
| GigaChat (Sber) | 8.0 | 7.5 | 7.0 | 9.0 | Medium | Yes |
| Qwen (open) | 8.0 | 8.0 | 8.5 | 6.0 | Low (self-host) | Locally |
GRAMMAR AND STYLE: WHERE IT BREAKS
On pure literacy the gap is minimal. Claude and GPT make almost no agreement errors even in long subordinate clauses. DeepSeek sometimes drops verb aspect in complex temporal constructions. YandexGPT is expectedly strong — the model has seen a lot of living Russian, not translated.
Interesting point: Western models write "correctly, but not in a Russian way". Stylistically it's slightly translated text — long periods, excess of deverbal nouns. YandexGPT and GigaChat sound more natural in conversational register, but sometimes lapse into officialese.
Typical trap
Ask to rewrite "Control over the implementation of measures is assigned to..." — frontier models give a literate but still bureaucratic variant. YandexGPT more often guesses that they want living language and cuts bureaucratese more aggressively.
DATA EXTRACTION AND CODE
Western models are confidently ahead here. Structured output, strict JSON by schema, tool calling — the ecosystem is polished. YandexGPT and GigaChat support function calling, but are more capricious: they break the schema more often, add extra fields, wrap JSON in a markdown block when you ask for a clean object.
Here is a minimal test run to check any model on your extraction task. Example for an OpenAI-compatible endpoint (provided by almost all, including proxies to DeepSeek and local Qwen via vLLM):
import json
from openai import OpenAI
client = OpenAI(base_url="https://api.example.com/v1", api_key="...")
PROMPT = """Извлеки данные из текста строго в JSON.
Схема: {"name": str, "city": str, "age": int|null}
Верни ТОЛЬКО валидный JSON без markdown.
Текст: Меня зовут Пётр, живу в Нижнем Новгороде, мне 34."""
resp = client.chat.completions.create(
model="your-model-id",
messages=[{"role": "user", "content": PROMPT}],
temperature=0,
)
raw = resp.choices[0].message.content.strip()
# частая беда рус-моделей — обёртка ```json ... ```
if raw.startswith("```"):
raw = raw.split("```")[1].removeprefix("json").strip()
data = json.loads(raw)
assert data["city"] == "Нижнем Новгороде" or data["city"] == "Нижний Новгород"
print(data)
Note the line cleaning markdown — this is not paranoia, but a real workaround still needed for GigaChat and some open models. Claude and GPT return clean JSON stably with an explicit instruction.
LOCAL KNOWLEDGE AND 152-FZ
Here the situation flips. Ask about the procedure for submitting documents to a specific agency, current rates, or regional realities — YandexGPT and GigaChat answer more substantively because they are fine-tuned on a relevant corpus. Western models either politely decline or hallucinate outdated data.
But the main argument for Russian models is often not quality, but compliance. If you have personal data and localization requirements (152-FZ), sending them to a foreign API is a legal risk. YandexGPT and GigaChat keep data in Russia and provide contracts for government procurement. For many enterprise projects, this decides the choice before any benchmarks.
- Maximum quality needed, data not sensitive — Claude or GPT.
- Personal data / government procurement — YandexGPT or GigaChat, no alternatives.
- Saving on volume — DeepSeek or self-host Qwen.
- Full privacy / offline — open model on own hardware.
PRICE AND TOKENIZATION OF RUSSIAN
Economics heavily depend on how many tokens the model spends on Cyrillic. You can estimate like this:
# bash: грубая оценка стоимости на русском тексте
# считаем, что рус-слово ~ 1.5 токена в современном токенайзере
WORDS=$(wc -w < article_ru.txt)
TOKENS=$(( WORDS * 3 / 2 ))
echo "Примерно $TOKENS входных токенов"
# при цене X$/1M токенов:
python3 -c "print(f'{$TOKENS/1_000_000 * 3.0:.4f} USD при 3\$/1M')"
DeepSeek is an order of magnitude cheaper than flagships with comparable Russian. Open Qwen on own GPU is almost free for inference, but you pay for hardware and DevOps. YandexGPT and GigaChat are mid-price segment, but with ruble payment and no currency hassle, which for a Russian legal entity is sometimes more important than the price itself.
CONCLUSION: WHAT TO TAKE IN 2026
There is no universal answer — there are three scenarios:
- Product for wide audience, quality matters most — take Claude or top GPT, they write in Russian almost flawlessly.
- Russian legal entity, personal data, integration with local services — YandexGPT or GigaChat cover compliance and know local context.
- Large volumes or full privacy — DeepSeek via API or Qwen on self-host give the best price/quality ratio.
The gap in Russian quality between camps has collapsed to fractions of a point over the last year. The real choice in 2026 is determined not by grammar, but by three things: where your data lies, how much you are willing to pay, and whether you need local knowledge.
Frequently Asked Questions
Which LLM understands Russian best in 2026?
In pure written Russian quality, frontier models Claude and GPT lead — minimum agreement and style errors. But in local knowledge and naturalness of conversational Russian, YandexGPT often does not concede, and surpasses in compliance.
YandexGPT or GigaChat — which to choose?
YandexGPT in my tests is more stable in grammar, summarization, and function calling. GigaChat is slightly weaker in structured output, but competitive in local knowledge and convenient if you are already in the Sber ecosystem. Both keep data in Russia.
Can foreign LLMs be used with personal data?
Legally risky due to localization requirements (152-FZ). If requests contain personal data of Russian citizens, YandexGPT, GigaChat, or self-hosting an open model on your own server in the Russian Federation is safer.
How many tokens does Russian text consume compared to English?
In modern 2026 tokenizers, a Russian word costs approximately 1.3-1.7 tokens versus ~1 for English. This is more expensive, but noticeably better than 2-3 tokens in previous-generation models.
I run new models in Russian as they are released and publish live runs, not press releases. If you want to catch fresh repositories and tools before everyone else — join the Telegram channel REDDYX AI — new repositories every 30-60 minutes.