How Many Tokens Your Task Will Eat

How the calculation works

The starting point is the volume of text the model must see in order to answer to the point. It is converted into tokens at a rate of 700 tokens per page. That rate follows from 1,800 characters on an A4 page and Cyrillic tokenization, where one token covers 2–3 characters.

Next, the volume is multiplied by request traffic. The three architectures differ in how much text goes into the model on each request.

Full text in the prompt. The entire base is inserted into every request. Consumption equals the base size multiplied by the number of requests. The repeating part is cached, and reading from cache costs less than full input — the calculation assumes a coefficient of 0.4 as an average across traffic where the cache does not always hit.

Retrieval over the base. Documents are cut into fragments, the question finds the matching ones, and only those go into the prompt. In the calculation that is 12 fragments plus instructions — about 5,200 tokens per request, regardless of base size. In exchange, a fixed line item appears: the vector store and its upkeep, 18,000 RUB a month.

Fine-tuning. Knowledge is moved into the model's weights, and the prompt stays short — about 900 tokens. Per-request consumption is minimal, but inference on a fine-tuned model costs roughly one and a half times more than on the base one. Plus two recurring bills: a training run from 15,000 RUB per cycle and dataset preparation at 900 RUB per page, spread across the year.

Output tokens are counted identically for all three options and priced at four times the input rate — that is how cloud provider price lists are built.

What moves the bill the most

First, the product of base size and traffic. In the first option these two numbers multiply directly, which is why it takes off faster than the others. Two hundred pages and three thousand requests give 420 million input tokens; the same pages at thirty thousand requests give 4.2 billion.

Second, the model class. The gap between the light and the top-tier model in the calculation is elevenfold. It stretches all three columns at once, so it usually does not change the winner, but it moves the conversation from tens of thousands to hundreds.

Third, how often the base is updated. For retrieval, re-indexing is nearly free: embeddings cost single-digit rubles per million tokens. For fine-tuning, every update means a new training cycle, and weekly edits multiply that line item fourfold.

The conditions under which each option wins

Full text in the prompt pays off while the base fits into the context window and traffic is measured in hundreds of requests. There is almost no development, and documents can be changed the same day — an edit to a file is immediately visible to the model.

Retrieval over the base pays off at medium and large traffic with a living knowledge base. The fixed 18,000 RUB dissolves at a few thousand requests already, and per-request consumption stops depending on whether the base grew to five hundred pages or to three thousand.

Fine-tuning pays off at high, uniform traffic with a stable base. A short prompt gives the lowest cost per request, and the one-off investment returns faster the less often retraining is required. Infrequent updates are the key condition here: a quarterly cycle is twice as cheap as a monthly one.

What stays out of scope

Development. The calculation shows operation — the bill that arrives every month after launch. Integration with email, CRM, and access rights is counted separately and depends little on the architecture.

The state of the documents. Policies stored as scans and instructions that contradict each other make any of the three paths more expensive, and for fine-tuning they double dataset preparation. That line item can only be estimated by looking at the files themselves.

Answer quality. The cheap option wins on money, not on hitting the substance of the question. The usual order is this: first a pilot on a hundred real questions, then a comparison of accuracy, and only then a choice by price among those that passed on quality.

How to use the result

It is worth looking at all three columns at once. The ratio between them holds even if the price list is off by a factor of two — and it is that ratio which determines where next year's money goes.

The «overpayment on the worst choice» line shows the price of a decision made out of habit. When it is measured in hundreds of thousands a year, an evening spent on this calculation comes back with room to spare.