How it worksWhy it worksPricingQuestionsBlog

How much does a chatbot cost? Here are the real numbers

A website chatbot costs between $0 and $199 a month on a self-serve plan. Underneath that, the AI cost of one answer is about half a cent: chatfor.site measures 1,794 input and 87 output tokens per answer, which at Claude Sonnet 5 rates is $0.00446. Everything above that is margin, hosting and people.

Almost nobody in this market publishes that second number. Vendors quote a monthly price and stop, which leaves you unable to tell whether $150 a month is a fair price or a 30 times markup. This article gives you both figures, shows the arithmetic, and then shows you how to run the same sum for your own shop. If you want the wider picture first, start with our ecommerce chatbot guide and come back for the money.

What actually goes into the price of a chatbot

A chatbot's cost is the sum of four things: the AI cost of generating each answer, the cost of indexing your documents so answers can be found, the fixed hosting and database bill, and the human time spent writing and correcting the source material. Only the first varies with traffic.

That split matters because vendors price on different units, and the unit is where the money hides:

  • Per seat. You pay for each human agent with a login, whether or not the AI answers anything.
  • Per resolution or outcome. You pay when the AI settles a question without a person.
  • Per conversation. You pay once per visitor thread, however many messages it contains.
  • Per message credit. You pay per exchange, and a credit is not always one answer.
  • Per answer. You pay for each question the assistant answers, which is how chatfor.site meters.

Two tools at the same headline price can differ by an order of magnitude once you translate their unit into the thing you actually care about: a customer question, answered.

Worth knowing: the unit is the negotiation. Ask any vendor to convert their price into cost per answered customer question at your volume. If they cannot, you are not being sold a price, you are being sold a range.

How much does a chatbot cost per answer?

Here is the whole calculation, with nothing hidden.

chatfor.site generates answers with Anthropic's Claude Sonnet 5, priced at $2 per million input tokens and $10 per million output tokens (Source: Anthropic documentation, verified 4 September 2026). Across a real support corpus, the measured averages are 1,794 input tokens and 87 output tokens per answer. Those figures live in lib/plans.ts as MEASURED_TOKENS_PER_ANSWER, not in a marketing deck.

input:   1,794 tokens / 1,000,000 x $2  = $0.003588
output:     87 tokens / 1,000,000 x $10 = $0.000870
                                  total = $0.004458

So one answer costs $0.00446 in AI, or 0.446 of a cent. A thousand answered questions cost $4.46. Ten thousand cost $44.58.

The input number is high relative to the output number because retrieval sends the model the relevant passages from your own documents on every question. That is the whole design: the assistant reads your delivery policy before it answers a delivery question. Input tokens are the price of that, and they came down from 2,373 to 1,794 through three changes to how the context is assembled, which is a 24 per cent reduction in the dominant cost line.

Two costs that people expect to see here are absent:

  1. Indexing is free of per-document API cost. Embeddings run locally through transformers.js with the bge-small-en-v1.5 model, so uploading a 40 page returns policy does not bill you a single token.
  2. A curated question-and-answer pair that matches above the verbatim gate never reaches the model at all. It costs nothing. A shop with a well-kept list of its twenty most common questions serves a meaningful share of its traffic for zero AI cost.

What the plans cost, and what they cost to serve

The table below is the honest version of a pricing page: what you pay, what the allowance is, what each answer works out at if you spend the whole allowance, and what that allowance costs in AI to serve.

Plan Price Answers a month Your cost per answer AI cost to serve the full allowance
Free $0 forever 100 $0 $0.45
Starter $29 a month 2,000 $0.0145 $8.92
Growth $79 a month 6,000 $0.0132 $26.76
Business $199 a month 20,000 $0.00995 $89.20

Read the last two columns together. On Starter you pay 3.3 times the AI cost of an answer; on Business you pay 2.2 times. The gap between the two funds hosting, the Postgres database, support, and the fact that most accounts never spend their full allowance. At full usage the gross margin is 69 per cent on Starter, 66 per cent on Growth and 55 per cent on Business.

Publishing that is uncomfortable and also the point. A price you can check is a price you can trust, and a shop owner deciding between four tools deserves to know which part of the bill is compute and which part is somebody's margin.

Worth knowing: the per-answer cost above is deliberately pessimistic. It bills every answer at the model rate, ignores the curated answers that never reach the model, and excludes nothing in your favour. Your real cost per answer will be lower than $0.00446, not higher.

Work out your own cost per answered question

You do not need a spreadsheet, you need last month's inbox. Five steps:

  1. Count the repeat questions. Open the last 30 days of email, Instagram DMs and contact form submissions. Count only the ones where the answer already exists in writing somewhere: delivery times, returns, sizing, materials, stock. Call that number Q.
  2. Multiply by the per-answer cost. Q x $0.00446 gives you the raw AI cost of answering all of them. For 300 questions that is $1.34.
  3. Divide the plan price by Q. At 300 questions on the $29 Starter plan, each answered question costs you $0.097. At 1,800 questions on the $79 Growth plan it is $0.044.
  4. Add the setup time. Gathering and uploading the documents is a one-off hour or two. Reviewing answers weekly is perhaps fifteen minutes. Price your own time honestly and amortise it over a year.
  5. Compare against doing nothing. This is the step people skip, and it is the only one that decides anything. See the next section.

If you want to run the token maths yourself against a different model, the formula generalises:

cost per answer = (input tokens x input price + output tokens x output price) / 1,000,000

Swap in the input and output token counts your own tool reports and the per-million prices from the model provider's pricing page. Any vendor that cannot tell you its average token counts per answer either has not measured them or would rather you did not do this sum.

Is a chatbot cheaper than answering the questions yourself?

Compare it to the two things it replaces, not to zero.

Against your own time. If 300 questions a month each take three minutes to read and reply to, that is 15 hours. At any plausible value for an owner's hour, $29 is not a close call. The comparison only turns when the volume is tiny, and at tiny volumes the free plan's 100 answers a month covers you anyway.

Against a lost sale. Baymard Institute puts the average documented cart abandonment rate at 70.22 per cent, calculated across 50 studies (Source: Baymard Institute, 2026). The stated reasons are dominated by things a document can answer: 40 per cent abandon because extra costs were too high, 20 per cent because delivery was too slow, 13 per cent because the returns policy was not satisfactory, and 12 per cent because they could not see or calculate the total order cost up front (Source: Baymard Institute, 2026). None of those are solved by a chatbot pretending to be a salesperson. Several of them are improved by a visitor being able to ask "how long to Manchester and what if it does not fit" and getting your actual policy back in four seconds.

One recovered order a month at a typical basket value pays for a year of Starter. That is the arithmetic that matters, and it is also why the honest answer to "how much does a chatbot cost" is a range rather than a number: it depends on what an unanswered question costs you.

What to check before you pay for anything

Cost is not only the invoice. Five things reliably turn a cheap tool into an expensive one:

  • Overage behaviour. When you pass the allowance, does the assistant stop answering or does the bill grow silently? Ask, and get the answer in writing.
  • Annual lock-in. A monthly price shown only on annual billing is a twelve month commitment wearing a monthly disguise.
  • Seat minimums. Per-seat pricing bills your team, not your customers, which is the wrong shape for a one-person shop.
  • What happens when the documents do not cover a question. A tool that guesses creates refunds, complaints and hours of cleanup, all of which are costs. chatfor.site answers only from what you uploaded and is instructed to say it does not have that to hand when the material does not cover the question.
  • Whether you can see why it said that. Being able to open a past conversation, see which passages an answer came from and overrule it is the difference between fixing a wrong answer in a minute and arguing about it for a week.

For the same axes applied to specific tools, see our ranked comparison of the tools on the market and, if you are currently quoted on per-seat pricing, our look at what else you can run instead of a per-seat helpdesk. If you have not yet decided whether you want automation at all, the case for a person versus a machine on each type of question is the better place to start.

Frequently asked questions

How much does a chatbot cost per month?

Self-serve website chatbots run from $0 to about $199 a month. chatfor.site is free for 100 answers a month, $29 for 2,000, $79 for 6,000 and $199 for 20,000. Enterprise platforms priced per agent seat or per resolution can reach several hundred a month before any AI usage is counted.

Is there a genuinely free chatbot for a website?

Yes. chatfor.site has a free plan with 100 answers a month, one assistant and 25 documents, with no card required. It carries a small chatfor.site badge in the widget. That allowance is enough to test whether the assistant answers your real customer questions properly before you spend anything.

What makes one chatbot much more expensive than another?

The billing unit, not the technology. Per-seat pricing charges for your staff rather than your customers. Per-resolution pricing charges each time the AI settles a question. Per-credit pricing can consume several credits for one exchange. Convert every quote into cost per answered customer question before comparing.

How much does the AI itself cost per answer?

About $0.00446 at chatfor.site's measured averages of 1,794 input and 87 output tokens, priced at Claude Sonnet 5's $2 per million input and $10 per million output tokens. That is under half a cent. Indexing your documents adds nothing, because embeddings run locally rather than through a paid API.

How many answers a month will a small shop actually use?

Count the repeat questions in your last 30 days of email and DMs, then double it, because a chat bubble gets asked things nobody would email about. Most single-location shops land in the hundreds rather than the thousands, which is why the 2,000 answer Starter plan exists.

Do I need to pay more for more documents?

Not on most plans. chatfor.site allows 25 documents on Free, 200 on Starter and unlimited on Growth and Business. Because embeddings are generated locally with transformers.js, adding documents carries no per-document API charge, so document limits are a plan boundary rather than a cost pass-through.

Sources

Try it on your own documents

Upload what you already have and ask it something. The free plan needs no card, and it will tell you when your documents do not cover a question.

Start free