HeyMark
Content

Blog

How much does an AI chatbot cost per month?

9 min read

An AI chatbot for WhatsApp or Instagram doesn’t have a single price, because what you pay each month is the sum of three layers: what the messaging platform charges per message, what the software or AI model that writes the replies charges, and the work of setting it up and keeping it current. That’s why two quotes can look so far apart: one includes the WhatsApp messages and the setup, and the other charges only for the software and leaves the rest to you.

On WhatsApp, Meta charges per delivered message, and replies from a third-party chatbot are service messages: the first 1,000 each month per phone number are free, and since October 1, 2026 the rest are charged. On Instagram, the messaging API documentation doesn’t mention any per-message charge. The AI model, if you pay for it directly, is billed by the token. We read every source on this page on November 20, 2026. Prices change, so open each link before you budget.

The three layers, side by side

Layer Who charges for it How it’s charged
Messaging on WhatsApp Meta Per delivered message, by message category and the recipient’s country
Messaging on Instagram Meta doesn’t publish a rate in its messaging API documentation That documentation doesn’t mention a per-message charge
Software or AI model The tool’s vendor, or the model provider if you connect it yourself A vendor plan, or input and output tokens
Setting it up and running it Your team, a developer or the vendor Hours, or a setup fee if the vendor charges one

When you compare two offers, put them in this table. The one that looks expensive may include the WhatsApp messages and the setup; the cheap one may bill those two layers separately or leave them to you.

Layer 1: what the messaging platform charges

WhatsApp: per delivered message

According to Meta’s WhatsApp pricing page for developers, Meta charges for messages that are delivered, not sent, and doesn’t charge for messages your customers send you. Three rules matter for a chatbot:

  • Your chatbot’s replies are service messages, even when an AI writes them. The documentation on non-template messages says a service message can be powered by a person or by a third-party AI solution, always inside the 24-hour window the customer opens. Each phone number gets 1,000 free delivered service messages a month, unused ones don’t roll over, and since October 1, 2026 the ones after that are charged.
  • If the chatbot is Meta Business Agent, Meta’s AI agent, the math changes. Its replies aren’t service messages: they’re charged by the token, at one global rate of $2.00 per million that covers both processing and delivery. Meta estimates a message typically uses 20,000 to 25,000 tokens, about 4 to 5 cents, and that it varies with complexity. A message is charged as service or as Business Agent, never both.
  • Anything the bot sends outside that window is something else. A reminder or a promotion to someone who hasn’t written to you in the last 24 hours goes out as a marketing or utility template, with its own rate per country.

The same page asks you to have a payment method on your messaging account so service messages aren’t interrupted. Rates by country, the free window for ads that click to WhatsApp and what changed on October 1 are in how WhatsApp Business API pricing works. What Meta’s agent does and what it costs inside the app are in what Meta Business Agent is and how much it costs.

Instagram: no published rate in the documentation

The Instagram messaging API documentation explains how a connected tool receives and answers a professional account’s messages, and it doesn’t mention any per-message charge. It doesn’t say they’re free either. Read it this way: Meta doesn’t publish a per-message rate there, and what you pay for on Instagram is the tool you connect.

What that documentation does set is when a chatbot can reply: after the person writes, inside a window that lasts 24 hours for most actions, and up to 7 days if the conversation started from an ad that clicks to Direct messages.

Layer 2: the software or model that writes the replies

You pay for this layer in one of two ways, and each changes what “price” means.

If you subscribe to a tool, you pay for its plan. Before comparing two, find out what unit each one charges by (contacts, conversations, AI replies or team seats) and what happens when you go over the limit. If your volume sits near the limit, a low-priced plan with a small cap can end up costing more than a higher-priced one with room to spare.

If you build it yourself on a model’s API, you pay for tokens: what the model reads (your instructions, the business information, the conversation and the new message) and what it writes. These are two prices from Anthropic’s API pricing page, in US dollars:

Model Reading (input) Writing (output)
Claude Haiku 5.5, for prompts up to 100,000 tokens $0.10 $0.50
Claude Sonnet 5.5 $2 $10

Two things move this layer more than the model you pick:

  • The model reads far more than it writes. For every reply it rereads the instructions, the business details and the conversation so far. A three-line reply can cost mostly for what it read.
  • What repeats can cost less. If you reuse the same block of instructions, Anthropic’s prompt caching charges those reads at a fraction of the input price: $0.01 per million on Claude Haiku 5.5, according to the same page.

Layer 3: setting it up and running it

This layer has no published rate, because it’s hours. Count them even if you do the work yourself.

  • Setting it up: connecting the messaging account, writing the information the bot needs (prices, hours, delivery times, return policies), deciding which conversations go to a person and testing it with real questions. If you code it, add the program and the server that receives the messages.
  • Running it: updating the details when they change, reading conversations to fix replies, handling the ones that were handed off and keeping up with each platform’s rule changes. On October 1, 2026, Meta resumed charging for WhatsApp service messages past 1,000 a month: a budget built before that date came up short.

A worked example (fictional)

A fictional case: the business, the volumes, the tokens per reply, the hours and the WhatsApp rate are made up. The Claude prices are the ones on Anthropic’s page. A plant shop builds its own WhatsApp chatbot on the Claude API. In one month it sends 2,000 replies, all inside the 24-hour window and to numbers in a single country. Each reply reads 3,000 tokens and writes 200. Assume a service message in that country costs $0.010. In real life, according to Meta’s documentation, a service message costs the same as a utility message in that market; the rates by country are in WhatsApp API pricing.

Line Math With Claude Haiku 5.5 With Claude Sonnet 5.5
What the model reads 2,000 × 3,000 = 6 million tokens 6 × $0.10 = $0.60 6 × $2 = $12.00
What the model writes 2,000 × 200 = 0.4 million tokens 0.4 × $0.50 = $0.20 0.4 × $10 = $4.00
WhatsApp service messages 2,000 delivered, 1,000 free 1,000 × $0.010 = $10.00, fictional $10.00, fictional
Total cash for the month $10.80, fictional $26.00, fictional
Setting it up 10 hours, once Your hours Your hours
Running it 2 hours a month Your hours Your hours

What the math shows:

  • The model is the line that swings most with your choice: $0.80 against $16, twenty times more. With the cheaper model, the WhatsApp messages and the hours weigh more than the AI.
  • Tokens per reply are the shakiest assumption. If every reply rereads a long conversation, the reading grows. Measure with your own conversations before you project a month.
  • What the example leaves out: the server the program runs on, any templates you send outside the window and customers with numbers in other countries, which have other rates.

With Meta Business Agent through the API, the same 2,000 replies would cost between $80 and $100 using Meta’s estimate (2,000 × 20,000 to 25,000 tokens, at $2.00 per million). That price already includes delivery, so you don’t add service messages, but there are no 1,000 free either: those replies aren’t service messages. In exchange, you don’t write or host the program that generates the replies. Loading the business information, deciding handoffs and reviewing conversations (layer 3) is still your work. The difference is what you pay to skip coding the bot.

What to ask a vendor before you pay

If you’re buying rather than building, these questions place each offer in the three layers:

  1. Are the WhatsApp messages included, or billed separately? Meta’s pricing page says a solution provider is charged only for messages sent through its own messaging account. Ask whether the account is yours, with Meta’s invoice in your name, or theirs, with the messages inside their price.
  2. What unit do they charge by, and what happens past the limit? Contacts, conversations, AI replies or team seats. Ask for the price at your volume, not the smallest plan’s.
  3. Do they charge for setup, and who updates the business information? If they do it, every change to your prices or hours depends on their time.
  4. How does a conversation get to a person? And can you review a reply before it goes out, or pause the bot in a single conversation?

If your messages come in on Instagram, Facebook or TikTok

HeyMark replies to comments and direct messages with AI in your brand’s voice, with approval and handoff to a person, on every plan. For what always gets the same answer, sending a DM to people who comment a keyword works on Instagram, Facebook and TikTok without using credits (on TikTok, the DM depends on what the platform allows in each country). It works with Instagram, TikTok, Facebook, LinkedIn and YouTube. WhatsApp isn’t one of them: for a WhatsApp chatbot, use the options above.

HeyMark has no free plan; every plan starts with a 7-day trial, and what each one costs is on HeyMark’s plans and pricing. How the AI replies is in AI comment replies, and the comparison between fixed replies, chatbot builders and AI for DMs is in AI chatbot for Instagram.

Plan your next post in HeyMark.

Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.