How it calculates
Input tokens per month: requests per day, multiplied by 30 and by the request length with context. Output tokens: requests multiplied by the answer length. The bill: input tokens multiplied by the price per million input tokens, plus output tokens at the output price, in US dollars. Russian models are priced per 1,000 tokens, with the same price for input and output, and are converted at the Bank of Russia rate of 84.4 per US dollar on 29 September 2026. The number of tokens depends on the model: per data from a large Russian bank, its model averages 3–4 characters per token, while other models give their own count for the same phrase. Measure a typical request with the provider’s token counter.
Prices in the calculator as of 29 September 2026
Foreign models, US dollars per million tokens, input and output: Claude Fable 5.1 10 and 50, Opus 5.5 4 and 20, Sonnet 5.5 2 and 10, Haiku 4.5 1 and 5, batch mode minus 50%. OpenAI GPT-6 Astra 10 and 50, GPT-5.5 5 and 30, GPT-6 Sol 2 and 10, GPT-5.4 mini 0.75 and 4.50, GPT-6 Luna 0.10 and 0.50. Gemini 3.1 Pro 2 and 12, and 4 and 18 above 200,000 tokens of context; Gemini 3.8 Flash 0.75 and 3.75 on a promotion until 31 December 2026; 3.1 Flash-Lite 0.25 and 1.50. Russian models, per million tokens: a large Russian tech company’s Pro model $14 in sync mode and $7 in async mode, its Lite model $2 and $1; a large Russian bank’s Max model $8, Pro $6, Lite about $0.80, half price in async mode; the Pro package of 50 million tokens for $300 costs the same as sync mode.
Example: 1,000 requests a day
A support agent: 1,000 requests a day, 1,500 input tokens with history and 500 output tokens, for a total of 45 million input and 15 million output tokens a month. The bill: GPT-6 Luna $12, Gemini 3.1 Flash-Lite $34, the bank’s Lite model $47, Gemini 3.8 Flash $92, GPT-5.4 mini $100, Haiku 4.5, $100, the tech company’s Lite model $150, Sonnet 5.5 and GPT-6 Sol $250 each, Gemini 3.1 Pro $250, the bank’s Pro model $350, the bank’s Max model $450, Opus 5.5, $500, GPT-5.5, $700, the tech company’s Pro model $900, Fable 5.1 and GPT-6 Astra $1,200 each. A hundredfold spread for one scenario: the choice of model matters more than the choice of provider.
What else affects the bill
Context caching lowers the price of the repeated part of a request at large foreign providers, and the calculator counts it as a discounted share of the input. Batch or async mode for overnight processing gives minus 50% at Claude and at both Russian providers; check the terms with other providers. Paying for foreign APIs from Russia: the services do not work directly, Russian addresses are not on the lists of supported regions, and the payment goes through an intermediary with a commission. Add it in the commission field.
How to read the result
Compare the bill with the result of the local AI calculator at the same volume. At 60 million tokens a month on a cheap model, the cloud costs $12 to $100, and by our calculation no configuration in the local AI calculator pays back within 36 months. On flagship models at $1,200 a month, a workstation with one card pays back in about 17 months if the local model solves the same task; that is a separate question you check with a pilot.