Every website chat tool answers the same question for you. What differs, and what almost nobody explains clearly before you sign up, is when the meter starts running. One vendor bills you for a seat whether anyone writes in or not. Another bills per resolved conversation. A third sells you a bundle of AI credits and then sells you more when they run out.
If you have ever opened a pricing page and closed it again without knowing what your bill would look like in March, this is for you. Below are the published rates of four assistants, quoted directly from their own pricing pages, and a plain explanation of who each model actually suits.
The four billing models, side by side
These are list prices as published on 23 August 2026. No estimates, no extrapolation, no "typical customer" math. Just what each page says.
| Caretta | $0.0355 per AI answer, from a prepaid balance. Plus $0.15 per page, once, to learn your site. | A visitor gets an answer | Crisp (Mini) | $45 per month, roughly $5 of AI credits included | The month begins | Tidio (Growth) | $49.17 per month, Lyro AI add-on from $32.50 per month | The month begins | Intercom (Essential) | $29 per seat per month, plus $0.99 per Fin outcome | The month begins, then again per resolution |
Read that third column again, because it is the whole story. Three of these four start charging on the first of the month regardless of what happens next. One starts charging when a visitor actually gets helped.
Why we are not showing you a monthly total
We could have drawn a bar chart with four monthly totals and put ours at the bottom. We decided not to, and the reason is worth saying out loud: to produce those totals we would have to invent your traffic, guess how many conversations the AI resolves versus hands off, and extrapolate add-on pricing well past the volumes any of these vendors publish. The result would look precise and be fiction.
What we can honestly give you is the arithmetic and your own numbers. Take your monthly visitor questions, multiply by the per-answer price, and compare that to the subscription you are considering. If you get 400 questions a month, Caretta runs about $14. If you get 4,000, about $142. The line is straight and there is no cliff in it.
The quiet month problem
Here is the scenario nobody puts on a pricing page. You install a chat assistant in November. December is slow. January is slower. Your traffic is seasonal, or your campaign ended, or you simply had a quiet quarter.
A monthly subscription does not care. It bills. Pay-as-you-go does care, because there is nothing to meter. This is not a trick of presentation, it is just what the two models do when usage goes to zero, and it cuts the other way too: a very high-volume site with a small team may well find a flat monthly fee cheaper than per-message billing. The right question is not which model is better. It is which one matches the shape of your traffic.
Where per-seat pricing bites
Seat-based pricing has a particular failure mode for small teams. You are charged per person who can open the inbox, which sounds fair until you realise the people who need occasional access are the ones you will end up not giving access to. Your warehouse lead who could answer a shipping question in four seconds does not get a seat, because a seat costs $29 or $85 or $132 a month.
The AI layer usually sits on top of that, priced separately. Intercom's Fin, for example, is $0.99 per outcome on top of the seat fee. That is a clean, honest unit and for a company handling expensive B2B support tickets it can be excellent value. For a shop answering "do you ship to Ireland" forty times a week, it is a different calculation entirely.
What we built instead, and why
Caretta bills per answer because that is the only unit a website owner can reason about without a spreadsheet. There is no subscription, no seat count, no bundle to run out of. You put money in a balance, answers draw it down, and the balance is yours until it is spent.
Three things follow from that decision, and they matter more than the headline rate:
Unsure answers are free
When the assistant is not confident in an answer, it says so instead of inventing something, and that answer is not charged. This is not generosity. If we billed for low-confidence output we would have a quiet incentive to let the model guess, and we would rather not have that incentive at all.
Repeated questions cost less
The fortieth "do you ship to Ireland" is served from your knowledge base at $0.0122 instead of $0.0355, because it does not need a fresh model call. Popular questions are the cheap ones, which is the opposite of how per-conversation pricing usually works.
You set a daily ceiling
A daily budget cap is part of the product, not an enterprise feature. When the day's spend hits your limit, answers pause until tomorrow. Combined with per-visitor rate limits and optional bot verification, this is what makes "pay per use" safe rather than frightening.
Training: the cost we could have hidden
Before Caretta can answer anything, it reads your site. That crawl costs $0.15 per page, once, and it is the one number a pay-as-you-go pitch is tempted to leave out. A 60-page site costs about $9 to train, plus a small quality reserve that is quoted up front and refunded if it goes unused. You see the total and approve it before a single page is fetched.
Retraining after you change your content is the same deal: you see the price, you approve it, it runs.
Installing it
The integration is one line of HTML, the same shape as an analytics tag, which means it works on any site regardless of what built it. If you are on WordPress or Wix there is a packaged version so you never touch code at all.
Developers can skip the widget and call the REST API directly with one Bearer key, which is the same assistant behind a JSON interface.
Questions people actually ask before switching
Can I try it without committing?
There is no subscription to cancel, so "trying it" means topping up a small balance and installing the snippet. If it does not work out, delete the line of HTML and the widget is gone. Nothing recurring is left behind.
What happens if the assistant gets something wrong?
Answers are grounded in your own pages, so wrong answers usually mean the page itself is out of date or ambiguous. Low-confidence responses are flagged and not billed, and you can correct the knowledge base directly from your dashboard.
Does it handle languages other than English?
It answers in 45 languages and detects the visitor's language automatically. The widget's own interface, the buttons and labels, ships in English, Turkish and German.
How do I know what I am spending?
Every answer is a line item against your balance, visible in the dashboard, with a usage and spend report available through the API as well.
How to decide in five minutes
Pull last month's numbers and answer three questions. How many visitor questions did you actually get? How many people need to see the inbox? Does your traffic swing seasonally?
Heavy, steady volume with a large support team usually favours a flat subscription. Lumpy or seasonal traffic, or a team of one or two, usually favours paying per answer. Anything in between is worth a month of real data before you commit to twelve.
If you want to run the numbers on your own traffic, the calculator on the Caretta Assistant page uses the exact same rates the dashboard charges. Current pricing for all providers changes, so check the linked pricing pages before making a decision on anything you read here.