TL;DR. Since 1 October 2026, Meta charges for the bot's replies too. The first 1,000 replies per phone number each month are free, then every reply costs the same as a utility message. For a business in Portugal that is USD 0.0171 (EUR 0.0142) per reply. A reply from a small AI model costs about USD 0.0002 in tokens, so on a cheap model Meta's fee is roughly 85 times the AI bill. A bot that sends 3,000 replies a month to Portuguese numbers pays Meta about EUR 28.40.
I'm Denys, a full-stack and AI engineer. I built an interactive prototype of a WhatsApp and Telegram chatbot on sample chats. It has a visual flow builder, AI answers from the company's own documents, lead scoring and a handoff to staff. Below is how such a bot is put together and what it costs to run, worked out from Meta's rate card and the model vendors' price pages, all read on 15 October 2026.
How much does a WhatsApp chatbot cost? What Meta charges
Meta's part of the bill is USD 0.0171 (EUR 0.0142) per reply to a Portuguese number once the first 1,000 replies a month are used, and a small AI model adds about USD 0.0002 per reply. Meta charges per message the business delivers, never for messages customers send. The price depends on the message type and on the country code of the customer's phone number, and it doesn't depend on where your business is.
- Service messages are the bot's free-form replies inside the 24-hour window that opens when a customer writes. Since 1 October 2026 the first 1,000 per phone number each month are free. After that each one costs the utility rate.
- Utility templates are updates such as booking confirmations and reminders. Since 1 October 2026 they are charged even inside the 24-hour window.
- Marketing templates are promotions, offers and anything else that isn't a transactional update. They are the expensive category and never get cheaper with volume.
- Free entry point. When a customer arrives through a Click-to-WhatsApp ad and you reply within 24 hours, messages in that conversation are free for up to 7 days.
Meta's rate card effective 1 October 2026, per delivered message:
| Customer's country | Marketing | Utility, authentication and service |
|---|---|---|
| Portugal (billed as "Rest of Western Europe") | USD 0.0592 (EUR 0.049) | USD 0.0171 (EUR 0.0142) |
| Brazil | USD 0.0625 (BRL 0.3217) | USD 0.0068 (BRL 0.035) |
| Spain | USD 0.0707 | USD 0.0200 |
| Germany | USD 0.1365 | USD 0.0550 |
| United Kingdom | USD 0.0635 | USD 0.0220 |
| United States and Canada | USD 0.0250 | USD 0.0034 |
Meta changes these rates only on the first day of a quarter. Check the official pricing page before you plan, because several popular guides still describe the old rules, where replies inside the window were free.
How much does each AI reply cost?
The model is billed in tokens. That's the text you send in (instructions, document passages, recent messages) and the text it writes back. A useful working assumption is about 1,600 tokens in and 80 tokens out per reply. One chatbot vendor reports that average from its own logs. Your bot will differ, so measure it once it runs.
Every business picks a different model, so here is the same reply priced on several. Prices are OpenRouter's list on 15 October 2026, which matches the vendors' own price pages:
| Model | Price per million tokens, in / out | One reply (calculated) | 3,000 replies a month |
|---|---|---|---|
| DeepSeek V4.1 Flash | USD 0.024 / 0.72 | about USD 0.0001 | USD 0.29 |
| Claude Haiku 5.5, GPT-6 Luna | USD 0.10 / 0.50 | about USD 0.0002 | USD 0.60 |
| Qwen3.8 Flash | USD 0.15 / 0.47 | about USD 0.0003 | USD 0.83 |
| Mistral Small 4 | USD 0.15 / 0.60 | about USD 0.0003 | USD 0.86 |
| Gemini 3.5 Flash Lite | USD 0.30 / 2.50 | about USD 0.0007 | USD 2.04 |
| Gemini 3.8 Flash | USD 0.75 / 3.75 | about USD 0.0015 | USD 4.50 |
| Claude Sonnet 5.5, GPT-6.1 Sol | USD 2 / 10 | about USD 0.004 | USD 12 |
| Claude Opus 5.5 | USD 4 / 20 | about USD 0.008 | USD 24 |
Sources: Meta rate card from 1 October 2026, OpenRouter list prices, 15 October 2026.
So the choice of model moves the AI bill about 80 times, from USD 0.29 to USD 24 for the same 3,000 replies. Even on Claude Opus 5.5 that is less than Meta's USD 34.20 for those replies in Portugal. On a small model, Meta's fee is the bigger line by far. A small model is often enough for answering questions from your own documents, and you can send only the hard cases to a larger one.
What an agent costs beyond its replies, from the build to hosting, is in how much an AI agent costs to run.
What does one bot cost per month in different countries?
Take one bot on one phone number that sends 3,000 replies a month, all inside the 24-hour window, on a small model. These are calculations from the prices above and not an invoice:
| Customers in | Meta: 2,000 paid replies | AI: 3,000 replies on a small model | Total per month |
|---|---|---|---|
| Portugal | USD 34.20 (EUR 28.40) | about USD 0.60 | about USD 35 |
| Brazil | USD 13.60 (BRL 70) | about USD 0.60 | about USD 14 |
| Spain | USD 40.00 | about USD 0.60 | about USD 41 |
| Germany | USD 110.00 | about USD 0.60 | about USD 111 |
| United Kingdom | USD 44.00 | about USD 0.60 | about USD 45 |
| United States | USD 6.80 | about USD 0.60 | about USD 7 |
Add hosting for the webhook and the knowledge base, and any partner fee if you connect through a WhatsApp solution provider instead of Meta's Cloud API directly. Meta itself charges no extra fee for using the platform directly. Partners set their own markup.
The same bot sending 1,000 marketing messages a month would add USD 59.20 in Portugal and USD 136.50 in Germany. Promotions are where WhatsApp gets expensive, and answers aren't.
How is a WhatsApp AI chatbot built?
Five parts, in the order a message travels:
- The channel. WhatsApp messages arrive through Meta's WhatsApp Business Platform as webhook calls to your server. The same bot can serve Telegram and a website chat widget. The prototype handles all three in one inbox.
- The flow. A visual flow builder decides how a conversation goes. The prototype has nine kinds of steps: start, message, question, AI response, condition, action, delay, handoff and end. Fixed steps handle the greeting, the qualifying questions and the handoff, and the AI fills only the open parts.
- The knowledge base. The bot answers from documents the company uploads and not from whatever the model remembers. Every AI reply carries the source it used, so staff can check it.
- The handoff. When the bot is unsure, it hands the chat to a person instead of guessing. Below a confidence line, a human takes over in the shared inbox.
- The lead record. Each contact is marked cold, warm, hot or customer, gets a short summary, and hot leads go to the CRM.
Why should the bot answer only from your documents?
A model left to answer freely sounds sure even when it's wrong, and nobody can check where an answer came from. In the prototype I chose answers from uploaded documents, with a handoff below a confidence line. The cost of that choice is that someone has to keep the documents current, or the bot will confidently repeat last year's prices.
Who should control how a conversation goes?
The team should, and the model shouldn't. When the model leads every chat, conversations drift and a lead can leave without being asked the questions that matter: what they need, when, and how to reach them. A flow builder keeps those steps fixed and lets the AI handle what customers actually type in between. The cost is that every flow needs an owner who updates it when the offer changes.
What can go wrong with the bill?
- No payment method on file. Without one, Meta still delivers the first 1,000 free replies each month, then stops delivering replies until you add one.
- Reminders sent as the wrong type. A message Meta treats as marketing costs about 3.5 times a utility message in Portugal and 9 times in Brazil. Keep reminders and confirmations strictly informational.
- Prompts that grow. Sending the whole chat history and every document with each reply multiplies the token bill quietly. Send the last few messages and the few passages that match the question.
- Old guides. Many pricing pages written before 1 October 2026 still say replies inside the window are free. For a bot that mostly answers, that's the line that changed.
Platform subscription or a custom bot? What providers charge
Meta's fee and the AI bill are only part of the price. A provider that builds or hosts the bot charges for that on top. These are the entry prices the providers publish, copied as shown on the source pages and read on 22 October 2026. Currencies differ because each provider lists its own.
| Provider | Entry plan | Includes | Checked | Source |
|---|---|---|---|---|
| SoyAgentia | USD 59 a month (Basico), USD 99 (Plus), USD 199 (Enterprise) | 1,200, 3,500 or 7,500 AI replies a month, USD 15 per extra user | 2026-10-22 | SoyAgentia |
| Wati | USD 69 a month (Growth, monthly billing, US page) | 3 users, 500 active contacts a month | 2026-10-22 | SoyAgentia comparison |
| ManyChat | USD 39 a month (Pro, or USD 29 billed yearly) | 2,500 active contacts, USD 0.05 per extra contact on monthly billing | 2026-10-22 | SoyAgentia comparison |
| Chately | USD 39 a month (Esencial), USD 89 (Crecimiento), USD 159 (Empresarial) | Flows, AI agents with monthly AI credits, shared inbox, CRM | 2026-10-22 | Chately |
| Subscription platforms in Spain | EUR 50 to 300 a month | Visual flow editor, and some add their own fee on top of Meta's | 2026-10-22 | NovaChat IA |
| Custom-built bot | EUR 1,500 to 3,500 once, then from EUR 150 a month | Flow design, knowledge base, integrations, maintenance | 2026-10-22 | NovaChat IA |
| Local setup package (Malaga) | From EUR 290, 490 or 790 once, then EUR 25 to 80 a month | FAQ bot, bot with booking, bot with CRM, hosting and upkeep | 2026-10-22 | Potencia Redes |
Every provider above sells its own product, and most of these pages are written by the vendors themselves, so read them as list prices and not as a neutral survey. The subscription doesn't include Meta's per-message fees or, on most plans, AI usage beyond the replies it counts. Both are explained earlier in this article. A custom bot is worth its setup cost when the flow has to connect to your own booking system or CRM, or when per-contact subscription pricing would grow faster than your traffic.
When is a chatbot the wrong choice?
When most of your customers write once a month with a question only a person can answer, a shared inbox with saved replies may do the job for less. A bot pays off when the same questions arrive every day, at hours when nobody is there to answer.
What these numbers do not tell you
The monthly figures are calculations from list prices and not an invoice. The prototype runs on sample chats. Token counts depend on your prompts and documents, so measure them in your first week. Prices are in USD as Meta and the model vendors list them, before VAT, and they change. Meta's change on the first day of a quarter, and the model vendors' change whenever they announce it.
Sources
- Meta, WhatsApp Business Platform pricing and its rate cards effective 1 October 2026 (USD, EUR, BRL), read 15 October 2026.
- OpenRouter, model list prices. Anthropic, Claude pricing. OpenAI, API pricing. Google, Gemini API pricing. All read 15 October 2026.
- Provider entry prices, read 22 October 2026: SoyAgentia (also the source for the Wati and ManyChat rows), Chately, NovaChat IA and Potencia Redes.
- Average tokens per reply: ReplyKit, How much does a WhatsApp AI agent cost, 5 September 2026.
Originally published at madsalt.tech. Written with AI assistance, checked by me.
Top comments (1)
Useful breakdown! The point that Meta’s fee dwarfs the AI bill on small models is something that most people don't consider.
I'm curious about the “send only the hard cases to a larger model” part. How do you decide this? When you switch models mid-conversation, how are you carrying the history across, given the providers’ messages often have different formats?