A WhatsApp chatbot's cost comes down to two bills. The build is paid once and grows with what the bot must do: answer, quote or book inside your system. The running bill has three meters: Meta's fees per delivered message, the AI model's tokens and the server. Each provider publishes its own prices, linked below.
Updated
How much does a WhatsApp chatbot cost? Five lines, often on different invoices
Search the question and you'll find ranges that don't agree, because they aren't pricing the same thing. One is a menu bot on a builder platform. Another is an agent that books inside your calendar. A third shows the subscription and leaves Meta's fees for later.
Every WhatsApp chatbot's cost splits into the same five lines. Two are work: building it and keeping it right. Three are meters that run while it talks: Meta, the AI model and the server. Each can come from a different company, so when a quote shows one number, ask which of the five it covers.
| Line | How it's paid | What makes it grow | Who bills it |
|---|---|---|---|
| Build | Once | What the bot must do, the systems it connects to, languages and rules | Your agency or your own developers |
| Upkeep | Whenever something changes | New flows, price and rule changes, new templates | Your agency or your own developers |
| Meta's WhatsApp fees | Per delivered message | Message category, the customer's country, monthly volume | Meta, or the platform you send through |
| AI model | Per token | Conversation length, instructions sent on every turn, the model you pick | OpenAI, Anthropic or another provider |
| Hosting | Monthly | Server size, database, backups | Your hosting provider |
How much does it cost to build a WhatsApp chatbot? It depends on the job
The price doesn't depend on WhatsApp: the channel is the same for every bot. What grows is the job you hand it. There are four steps, listed below, and each one adds work to the one before.
Then come the multipliers. Every system the bot talks to (calendar, booking panel, store, CRM) is one more integration. Every language needs its own tone and its own testing. Every rule it can't break needs a written answer, and so does the moment a person takes over. Most of a build goes into those edge cases, not into the conversation that goes right the first time.
At Al Teu Aire, the agent sits on the fourth step. It books events for more than 40 venues in Barcelona, reads availability and prices live from the team's panel and answers in Spanish and Catalan. Not sure which step you need? Our guide on what to automate first with AI scores the candidates.
- Answers from your information. Hours, services, policies, directions. The smallest build, and sometimes all you need.
- Qualifies and hands over. Asks the questions your team would ask and drops the lead into your CRM with the chat attached.
- Quotes live. Reads prices and availability from your system the moment the customer asks, so it never quotes last month's rate.
- Books or orders. Writes the booking or the order into your system and sends the confirmation.
WhatsApp Business API pricing in 2026: what Meta bills and what it doesn't
Meta doesn't charge for the chatbot. It charges for the messages your business sends through the WhatsApp Business Platform, and only once they're delivered. Since July 1, 2025 it bills per message, not per 24-hour conversation, so a guide that still prices “conversations” predates the switch.
Each charged message is priced by its category and by the customer's country, read from the phone number's country code. Templates (marketing, utility and authentication) can go out at any time to customers who agreed to hear from you. Service messages are the free-form replies your bot writes while the customer service window is open: 24 hours from the customer's last message or call.
October 1, 2026 changed the math for AI agents. Each business phone number now gets 1,000 free service messages a month, and every one after that is billed. Utility templates sent inside an open window, free since July 2025, are now billed as well. A busy agent that used to reply for nothing now runs on a meter. One practical consequence: a reply split into three short messages counts three times.
All of this comes from Meta's pricing page for the WhatsApp Business Platform, checked on November 2, 2026. The rate for each country is in Meta's rate table on the WhatsApp Business website, also checked on November 2, 2026: pick the market, the currency and the message category, and it shows the rate per message.
| Message | When it can go out | Billed by Meta? (checked Nov 2, 2026) |
|---|---|---|
| Your customer writes to you | Any time | Never |
| The bot's free-form reply (service message) | Only while the 24-hour window is open | The first 1,000 a month per number are free, then per message |
| Booking confirmation or reminder (utility template) | Any time | Per message, also inside the window since October 1, 2026 |
| Offer or win-back message (marketing template) | Any time | Per message |
| One-time login code (authentication template) | Any time | Per message |
| Anything you send after a click-to-WhatsApp ad | Free entry point window, up to 7 days, if you reply within 24 hours | Free |
WhatsApp chatbot pricing for the AI: the model bills per token
The language model behind an AI agent is billed separately, by tokens: the small chunks of text it reads and writes. OpenAI and Anthropic both price them per million, with input and output counted apart and output costing more. Every model has its own rate. Both pages checked on November 2, 2026.
Input is the part people underestimate. The model doesn't remember the chat on its own, so on every turn the agent sends it the instructions, the conversation so far, the tools it can use and whatever data it looked up. A two-line reply can carry pages of input behind it.
One exception: Meta Business Agent, a category Meta added on July 1, 2026 for replies written by Meta's own AI, is billed by Meta per token. An agent that runs on your own code and model, like the ones we build, pays the model provider instead.
Keeping the model's bill down is a design job, not a negotiation:
- Cache the fixed part. Both providers bill cached input at a lower rate, so instructions that repeat on every turn get cheaper after the first.
- Fetch only what the question needs. The date the customer asked about, not the whole calendar. A price list pasted into the prompt is paid for on every message, and goes stale the day a rate changes.
- Trim the history. Turns that no longer matter can be summarized or dropped.
- Match the model to the step. A smaller model can sort an incoming message; the stronger one handles the conversation.
- Don't count on batch discounts. Both platforms charge less for batch processing, but that's for work that can wait. A customer on WhatsApp can't.
Hosting and upkeep: the lines a quote tends to leave out
A custom agent is software, and software runs somewhere. The server holds the agent's code, a database for conversations and bookings, logs you can read when an answer looks off, and backups. Your hosting provider bills it monthly, and it grows with traffic and storage, not with WhatsApp.
For reference, the agent at Al Teu Aire is built with TypeScript, PostgreSQL and Docker, alongside the WhatsApp Cloud API, OpenAI and Chatwoot. The code and the database live on a server. The WhatsApp Cloud API is hosted by Meta and the model runs on OpenAI's side, which is why those two show up as meters, not as hosting.
Upkeep is the line that gets forgotten. Prices change, a venue closes, a new rule comes in, a new template needs Meta's approval, and the agent has to know before the next customer asks. Meta's own rates and rules move on their own calendar too.
That's why our process doesn't stop at launch. The last step is reading real conversations, fixing the spots where the agent hesitates and adjusting the rules. How we build and run these agents is on our WhatsApp AI agents page.
How to estimate your WhatsApp chatbot's running cost before you ask for a quote
You can size the meters yourself in an afternoon. Pull last month's chats, answer the six questions below, and take the numbers to Meta's rate card and your model's pricing page. The first four size the meters. The last two size the build.
Bring the answers to a first call and the quote will be about your bot, not a generic one.
- How many conversations you get a month, and how many messages your side sends in each.
- How many of them start from a click-to-WhatsApp ad, and how many don't.
- Which messages you'll send unprompted (confirmations, reminders, offers) and to which countries.
- How many phone numbers you'll run. Meta's free service tier is per number.
- Where the facts live: calendar, booking panel, store, CRM or a spreadsheet.
- Which languages you need, and at which hours nobody is there to answer.
Work we have shipped
- Al Teu Aire
WhatsApp agent that books events for more than 40 venues in Barcelona. It reads availability and prices live from the team's panel instead of a pasted price list, and answers in Spanish and Catalan.
See the project
Frequently asked questions
Is a WhatsApp chatbot free?
Not really. What's free is the WhatsApp Business app's greeting and away messages, and those are saved replies: they can't read your calendar, quote or book. A real chatbot runs on the API, where the model and the server always cost money and Meta's free tier is limited (see the table above).
Do customers pay to message a WhatsApp chatbot?
No. Meta doesn't charge for messages customers send to a business. Every fee lands on the business side, and only for the messages it sends that get delivered.
Does it cost more if my customers are in another country?
It can. Meta prices each charged message by the country code of the number receiving it, so a Miami business writing to customers in Mexico or Colombia pays those markets' rates, not the US one. The currency you're billed in is picked when the WhatsApp Business account is created, from the 16 Meta supports (US dollars and euros among them), and it can't be switched freely later.
How often does Meta change WhatsApp Business API pricing?
Four times a year at most. Meta only updates pricing on January 1, April 1, July 1 and October 1, with at least a month's notice for new rates and six months for a new pricing model. Check the rate card every quarter, or have whoever runs your bot do it.
Is a chatbot builder platform cheaper than a custom agent?
Up front, usually: a subscription instead of a build, and it works well for menus and simple answers. Ask whether the plan includes Meta's fees or bills them on top. The limits show when the bot has to read and write in your systems. A custom agent costs more at the start, and it's your own software, not a rented one. If you'd rather wire it with an automation tool, see what n8n, Make and Zapier each bill.
Why doesn't Kivo Marketing publish a price?
Because two agents that look the same on WhatsApp can be very different builds: one answers questions, the other books inside your system in two languages. A published price would be wrong for almost all of them. We quote after the brief, once we know which step your bot sits on and which lines apply to you.