"How much does a chatbot cost" has two completely different answers, because two different things go by the same name. One is free and dim. The other charges per conversation and understands what you asked. Here are both, with August 2026 prices.
The free option: a script on your server
A bot that runs on rules. It matches keywords and returns a pre-written answer, offers buttons — "Prices", "Delivery", "Contact" — and walks the visitor down a tree of questions.
Cost: zero. It runs on hosting you already pay for, with no subscription and no per-conversation billing. You pay for the build, once.
What you get: answers to the twenty most common questions, contact capture, opening hours and delivery terms — instantly and reliably.
Where it breaks: the moment someone phrases things differently than you predicted. "Do you ship to Novi Sad?" passes; "will it get to NS by Friday" does not. The bot says "I did not understand" and the visitor leaves.
The paid option: a bot on an AI model
This one understands the question however it is phrased, reads your catalogue and terms, and composes an answer. You pay per use — per token, where a token is roughly three quarters of a word. You are billed for what the bot reads (input) and what it writes (output).
Price per million tokens
Two serious options for this job are Anthropic's Claude and xAI's Grok. Official rates, August 2026:
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Haiku 4.5 | $1.00 | — | $5.00 |
| Claude Sonnet 5 | $2.00 | — | $10.00 |
| Grok Build 0.1 | $1.00 | $0.20 | $2.00 |
| Grok 4.6 | $2.00 | $0.50 | $6.00 |
Both platforms bill repeated context more cheaply when it is cached — and for a bot that carries the same instructions and catalogue every time, that is the single biggest saving.
What that means per conversation
A realistic conversation: your instructions plus the slice of catalogue the bot must see (about 2,000 input tokens) and an answer of a hundred words or so (about 150 output tokens). Without caching, per thousand conversations:
| Model | Per conversation | 1,000 conversations |
|---|---|---|
| Server script | $0 | $0 |
| Grok Build 0.1 | $0.0023 | ~$2.30 |
| Claude Haiku 4.5 | $0.0028 | ~$2.75 |
| Grok 4.6 | $0.0049 | ~$4.90 |
| Claude Sonnet 5 | $0.0055 | ~$5.50 |
What this means in practice: a site with 200 conversations a month pays under one euro, whichever of the four you pick. The gap between cheapest and dearest at that volume is about fifty cents a month — so price is not the deciding factor. Cost only becomes worth thinking about at tens of thousands of conversations, and a company with that volume already has bigger support costs.
What the price difference buys
Since cost does not decide, behaviour does. What actually differs:
- How much the model can hold at once. Haiku 4.5 works with 200 thousand tokens, Grok Build with 256 thousand, Grok 4.6 with 500 thousand, Claude Sonnet 5 with a million. For a bot with a large catalogue that is the difference between seeing the whole range and seeing part of it and guessing.
- Following instructions. The most important trait for a business bot is not eloquence but how consistently it obeys the rules — especially "if you do not know, offer contact". Stronger models are generally more disciplined here.
- Connecting to systems. If the bot needs to check stock or an order status, it has to call your system. Both platforms support this; they differ in how reliably they pick the right call.
- Your language. Do not take anyone's word here, ours included. Take twenty of your real questions from the inbox and run them through both. The difference shows in half an hour and costs less than a dollar.
Our approach is simple: start with the cheaper model, since it handles most questions perfectly well. If testing shows it misses or ignores the rules, move up — the difference on the bill is too small to economise on.
Where the real costs hide
The token price is the smallest part of the bill. What the price list does not show:
- Preparing the data. The bot is only as accurate as the catalogue, price list and terms you give it. If those are a mess, they get fixed first.
- Testing on real questions. Not invented ones — the ones that actually arrive in your inbox.
- Maintenance. The offer changes, the answers change. A bot nobody looks at starts lying about prices within six months.
- Supervision in the first weeks. Someone has to read the conversations and catch where the bot missed.
How to choose
A server script if you have few products, questions repeat in the same shape, and the budget really is zero. It works — just do not expect understanding.
An AI bot if you have a broad range, customers ask in their own words, or you lose inquiries because nobody answers in the evening. Since the cost is a few dollars per thousand conversations, the decision is not financial — it is whether your data is tidy enough to feed it.
A combination is usually best: buttons for the common things, AI for the rest. Fast where it can be fast, smart where it has to be — and the bill stays small because the simplest questions never reach the model.
One rule that is not negotiable
The bot must be able to say "I do not know". A model with no answer in your data will, if allowed, produce something that sounds right. For a business an invented price or deadline costs more than a missing answer — because the customer arrives expecting what you promised. Configuring the bot to offer contact instead of a guess is not a detail; it is the condition.
We build chatbots and digital assistants — both scripts and AI, depending on what actually pays off.
Related services and pages
More from the blog
- What does a digital marketing agency do? (clear and concrete)
- Digital marketing for small businesses: where to start
- How to choose a digital marketing agency (question checklist)
- How much does Google Ads cost in Serbia?
Need a partner, not another PDF?
Rezident works as an external marketing team for SMEs — Google Ads, Meta, SEO, web and automation.
Book a call