Skip to content

How much does a conversation with an AI assistant cost? Tokens explained

Input tokens, output tokens, context and history: what really makes up the cost of a conversation with an AI assistant, and how to bring that cost down.

Costs and AI models5 min read

AI providers do not charge per conversation, per message or per month. They charge per token. The word appears in every price list but is rarely explained, which makes it hard to predict what an AI assistant will cost on a client’s site, let alone explain it to the client.

This article explains what a token is, what consumes tokens during a conversation, and how blabla.bot turns that consumption into the amount debited. It ends with a few practical ways to cut the cost without hurting answer quality.

What is a token?

An AI model does not read text letter by letter or word by word. It splits it into small pieces called tokens. A token can be a short word, part of a word, a punctuation mark or a space.

As a rule of thumb, OpenAI says one token is roughly 4 characters of English text. It is an average: the split depends on the model and the language. French text, for instance, usually takes a few more tokens than the same content in English.

Keep the order of magnitude in mind: a 4,000-character web page is about 1,000 tokens.

Input and output: two meters

Every call to an AI model has two meters.

  • Input tokens: everything sent to the model so it can answer.
  • Output tokens: the answer the model writes.

Both are billed, at different prices. With the providers in the blabla.bot catalogue, output tokens cost several times more than input tokens.

What goes into the input of an answer

This is where most of the cost hides. When a visitor asks a ten-word question, the model receives far more than those ten words. It gets everything it needs to answer well:

  1. The agent’s instructions: its role, mission, tone and forbidden topics. They are sent with every answer.
  2. Knowledge passages: the excerpts from the client’s pages and PDFs that best match the question. This is the principle of RAG (retrieval-augmented generation).
  3. Conversation history: the previous messages, so the agent understands a follow-up like “and on Saturdays?”. blabla.bot sends a limited number of them.
  4. The visitor’s question.

The question itself is often less than 1% of the input. Instructions and passages make up most of it.

A worked conversation

To estimate costs, the blabla.bot catalogue uses a reference conversation of 4 exchanges. In Simple mode, each exchange has about 4,000 input tokens and 250 output tokens. Here is the calculation for GPT-6 Luna · Simple, using the catalogue prices checked on 13 November 2026 (€0.10 per million input tokens, €0.50 per million output tokens, converted at 1 dollar to 1 euro).

StepCalculationAmount
Input of one exchange4,000 tokens × €0.10 / 1,000,000€0.000400
Output of one exchange250 tokens × €0.50 / 1,000,000€0.000125
Provider cost of one exchange€0.000525
Amount debited (×2)€0.001050
4-exchange conversation4 × €0.001050€0.0042

Under these assumptions, a conversation costs less than half a cent, and a thousand conversations cost about €4.20. Notice that input weighs over three times more than output, even though its unit price is five times lower: that is the effect of the amount of context sent.

These figures are calculation assumptions, not measurements. Long instructions, many passages or a drawn-out conversation increase the cost.

How blabla.bot counts

The method is the same for every answer:

  • the tokens actually consumed, input and output, are multiplied by the catalogue prices;
  • the result is multiplied by two, blabla.bot’s multiplier;
  • the amount is rounded up to the micro-euro (a millionth of a euro), once per answer.

Finding passages also has a small cost: the visitor’s question is turned into a vector to retrieve the relevant excerpts. That tiny cost is added to the answer’s cost before rounding.

All amounts are stored as whole micro-euros. Each answer records the prices applied when it was produced: if a provider changes its rates, your history stays the same.

Cutting the cost without hurting the answers

Since input weighs the most, that is where to act.

  • Write concise instructions. Ten clear lines beat a page of recommendations, and they are sent with every answer.
  • Look after the knowledge. Clear pages without duplicates or outdated content produce more relevant passages: better answers for the same number of tokens.
  • Pick the right mode. In-depth mode sends more passages and allows longer answers. It costs roughly twice as much as Simple mode. Keep it for sites whose questions justify it.
  • Set a monthly budget per site. It does not lower the cost of a conversation, but it caps a site’s total spend.

What to tell your client

Your client does not need to know about tokens. They need an order of magnitude and a guarantee. Talk in conversations per month and offer a package that includes a volume. On your side, the console gives you each site’s real consumption, and the monthly budget protects you from overruns.

To turn these assumptions into a monthly estimate for your own sites, use the calculator on the Pricing page.

Sources

Tags

  • tokens
  • costs
  • AI models