Loading prices...
All news
Glossy word-shaped bar cut into seven colourful rounded tiles, one tile dropping onto a stack of gold coins under an amber glow, dark steel blue background with circuit traces

Tokens in AI: why GPT-6.2 is priced per million and Russian text costs more

02:00 · 19.11.2026
Source: OpenAI tiktoken
2

OpenAI prices GPT-6.2 per million tokens: $10 of output for Sol and $40 for Astra. A token is the chunk of text a model reads and writes, usually a piece of a word. To see what that means for a real request, we ran our own news through OpenAI's tokenizer.

What we measured

We took this morning's story on MiCA licences, which we publish in English and Russian, and counted its body text with tiktoken, OpenAI's open-source tokenizer library. We used o200k_base, the encoding OpenAI introduced with GPT-4o. OpenAI has not said which tokenizer GPT-6.2 uses, so treat these as estimates.

  • English version: 469 words, 655 tokens, about 1.4 tokens per word
  • Russian version: 421 words, 775 tokens, about 1.8 tokens per word
  • the same Russian text on cl100k_base, the older GPT-4 encoding: 1,238 tokens

The Russian text has fewer words and still costs 18% more tokens. Single words show why: "Bitcoin" is one token, while the Russian spelling in Cyrillic splits into three tokens and the Ukrainian one into four. "Stablecoin" takes two tokens and its Russian spelling takes six.

What a request costs

At Sol's $10 per million output tokens, an answer the length of our English story costs about $0.0066, and the Russian one about $0.0078. On Astra at $40, the same answers cost about $0.026 and $0.031. Input is cheaper: $2 per million tokens on Sol and $8 on Astra.

Tokens also set the limit on how much a model can hold at once. A context window is counted in tokens too, so a long Russian document fills it sooner than the same document in English.

Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, opened his 2024 lecture on tokenizers with a warning.

“Tokenization is at the heart of a lot of weirdness in LLMs and I would advise that you do not brush it off.”

— Andrej Karpathy, AI researcher, from his lecture notes on tokenization

Quote source: Karpathy's minbpe lecture notes on GitHub.

In the same list of examples he names tokenization as the reason models do worse in non-English languages such as Japanese.

Our take

Maks, the Intokened editor, says the hardest AI term for him to grasp was LLM, the large language model itself. Tokens are the easier way in: the model reads and writes text in these chunks, and you pay for each one.

If you work in Russian or Ukrainian, budget about a fifth more tokens than for the same text in English, and test your own prompts in a tokenizer before you estimate a monthly bill. Our AI glossary explains the other terms you meet on a pricing page.

Token counts come from tiktoken with the o200k_base and cl100k_base encodings, run on 18 November 2026. Prices come from OpenAI's GPT-6.2 announcement of 17 November.

Published: 02:00 · 19.11.2026
Maks

Author

Maks

Trading man

I've been interested in the cryptocurrency market for a long time, am a trader, and write articles and news about my experience and crypto in simple terms.

Comments (0)

No comments yet — be the first!

Join our community

No spam, unsubscribe in one click