Meta Business Agent: the per-token price runs six times a utility message
By José Pedro · Co-founder of CRPRO
Meta Business Agent bills prompt and reply together at US$2.00 per million tokens, and a message typically eats 20,000-25,000 tokens — 4 to 5 US cents per interaction, delivery included. See the math, the break-even point against a flat CRM plan, and what to measure before deciding.
The Meta Business Agent is the AI agent Meta offers inside WhatsApp Business, and since August 1, 2026 it's been billed per token, not per message: US$2.00 per million tokens (price in effect as of August 2026), counting both the prompt and the reply. A message typically consumes 20,000 to 25,000 tokens, which works out to 4 to 5 US cents per interaction — somewhere between R$0.21 and R$0.26 at the August 2026 exchange rate. At that same rate, a utility message in Brazil costs about R$0.035. The AI's reply, then, runs six to seven times the price of a regular message. That amount already includes delivery: according to Meta, a message from the agent pays no service fee beyond the tokens.
Updated on October 31, 2026: the first version said tokens and message fees added up. Meta's documentation says otherwise, and we redid the break-even math.
That doesn't make Meta's agent expensive in every scenario. It makes its cost proportional to the number of conversation turns — which is exactly the variable nobody tracks.
How does the Meta Business Agent's per-token billing work?
There are three parameters, and the third one is what gets people.
| Parameter | Value (verified as of August 2026) |
|---|---|
| Price | US$2.00 per million tokens |
| Consumption per message | 20,000 to 25,000 tokens (typical value) |
| What's included in the count | Prompt and reply |
| Cost per interaction | ~US$0.04 to US$0.05 (~R$0.21 to R$0.26) |
| Billing start date | August 1, 2026 |
"Prompt and reply" means you pay for both sides of the call. The text the customer typed is the smallest part of that. The rest is what your business stacked on top: agent instructions, tone rules, a product table, exchange policies, the conversation history up to that point.
Since the price is per token and not per conversation, the invoice doesn't respond to the number of customers served. It responds to the number of times the agent got called — and the size of the payload traveling with each call.
Why does a two-line message consume 20,000 tokens?
A question like "do you deliver to Curitiba?" is under 10 tokens. The other ~20,000 come from the context that rides along with the question on every single call.
And here's the detail that changes the budget: the history gets resent on every turn. On the third bubble of the conversation, the two previous ones travel along again, paid for again. The cost doesn't scale with how much the customer writes. It scales with how much context you're carrying and how many times the customer replies.
That flips the intuition of whoever is building the budget: trimming the AI's replies helps very little, because the reply is the small slice of the bill; trimming the fixed prompt and limiting how many previous turns ride along helps a lot. A catalog pasted inside the agent's instructions is a recurring cost disguised as configuration — it gets charged again on every single message, forever.
How much does an AI interaction cost next to a utility message?
The implicit exchange rate in the rates in effect as of August 2026 is around R$5.10 per dollar. With that, the comparison looks like this:
| Item | Unit cost (US$) | In reais (approx.) | How many times a utility message |
|---|---|---|---|
| Utility or authentication message | ~0.0068 | ~R$0.035 | 1× |
| Marketing message | ~0.0625 | ~R$0.32 | ~9× |
| One Meta Business Agent interaction | 0.04 to 0.05 | R$0.21 to R$0.26 | 6× to 7× |
All the figures above are rates in effect as of August 2026, for the Brazil destination. The one-line takeaway: the AI's reply costs almost as much as a marketing message — the most expensive category in Meta's table — and that charge already includes message delivery.
The two caveats this whole post carries
The first is about the quality of each number. The marketing rate is the best-corroborated of the three. Utility and authentication rest on a medium-confidence estimate — write "about" on your spreadsheet. Even so, the conclusion holds: if utility ends up costing double what we estimated here, the AI interaction still runs around three times the message. The order of magnitude doesn't flip sides.
The second is about what each charge covers. Meta describes the token charge as a single charge that includes AI processing and delivery, and says a non-template message is charged either as a Meta Business Agent message or as a service message, never both. Anyone answering with third-party AI inside the 24-hour window pays the service message to Meta and the model cost to the AI provider.
An interaction isn't a conversation — and that's where the number grows
Every AI spreadsheet I've seen counts conversations. Meta counts interactions. A typical pre-sales conversation — greeting, qualifying question, reply about timing, reply about price, objection, close — runs 6 to 8 AI turns. The right math isn't "how many customers do I serve per month," it's "how many times does the agent open its mouth per month."
At R$0.21 to R$0.26 per interaction (August 2026 exchange rate and price), the month looks like this:
| Monthly interaction volume | Token cost alone |
|---|---|
| 1,000 | R$210 to R$260 |
| 5,000 | R$1,050 to R$1,300 |
| 10,000 | R$2,100 to R$2,600 |
| 30,000 | R$6,300 to R$7,800 |
Translating the middle row: 10,000 interactions is roughly 1,400 seven-turn conversations, a mid-size operation. That's R$2,100 to R$2,600 a month on the AI line, delivery included, before any platform.
Notice too that the per-interaction price is the same at any volume. There's no announced volume discount for token consumption, unlike what happens with utility and authentication on the message fee.
From how many interactions does AI bundled in the CRM get cheaper?
When automation and the agent run inside the platform you already pay for, the AI cost sits in the monthly fee, within the plan's AI conversation allowance. On CRPRO, the plan with the agent included costs R$499.90 a month, with 3 WhatsApp numbers and unlimited agents (price in effect as of August 2026). The difference is that here each reply is a Meta service message: every number gets 1,000 free per month and pays about R$0.035 per message beyond that.
Divide one by the other and the crossover shows up:
| Scenario | Tokens (Meta Business Agent) | Monthly fee with AI included plus Meta service messages |
|---|---|---|
| 500 interactions/month | R$105 to R$130 | R$499.90 |
| 2,000 interactions/month | R$420 to R$520 | about R$535 |
| 10,000 interactions/month | R$2,100 to R$2,600 | about R$815 |
The break-even lands somewhere between 2,100 and 2,700 interactions a month — 300 to 390 complete seven-turn conversations on a single number. Below that, the per-token model is cheaper and requires no contract at all. Above it, the gap grows in a straight line and doesn't stop growing, because one side is variable and the other isn't.
Treat that R$499.90 for what it is: a monthly fee that also pays for a shared inbox, a Kanban pipeline, campaigns, and reporting. The clean "AI versus AI" comparison is even more favorable to the flat fee than the table shows.
What changed with October 2026's charge?
Since October 1, 2026, Meta charges per message for service messages and utility templates sent inside the 24-hour window. We broke down the impact in who feels October's charge the most.
For anyone using AI, the effect depends on who replies. With the Meta Business Agent, the message is charged by tokens only. With third-party AI, the reply is a service message, and that is exactly the category with no volume discount.
The service-message rate is in Meta's current price table. If you find an article with a service-message figure that doesn't cite that table, it's a guess. The per-token price has been live since August 1, 2026.
What can be said: the right comparison sets the Meta agent's interaction against the service message plus the cost of the AI you use. Tokens and the service fee never fall on the same message.
When is the Meta Business Agent the right pick?
To be honest, there are scenarios where paying per token wins — and others where it loses badly.
| Operation scenario | Cheaper model | Why |
|---|---|---|
| Under ~2,000 interactions/month | Tokens (Meta Business Agent) | Below the break-even point, pay-per-use costs less than any monthly fee |
| Highly seasonal volume | Tokens | The variable cost follows the February slump; a flat fee doesn't |
| Early AI testing | Tokens | It's the cheapest way to find out whether AI handles your kind of question |
| High, steady volume | AI bundled in the platform | Above the break-even, the variable line just keeps growing |
| Long conversations with many turns | AI bundled in the platform | Every turn is a fresh charge, with the whole history riding along |
| A large prompt, with catalog and policies | AI bundled in the platform | The fixed context gets resent and paid on every message |
The practical cutoff is this: if your operation is small or irregular, tokens are the cheap route. If it's already steady and you already pay for a platform capable of running its own agent, the flat fee tends to come out ahead.
What to measure before deciding
Don't decide by a table — decide by your own numbers. Three measurements that fit into an afternoon:
- Count turns, not conversations. Pull 50 recent conversations and count how many messages went out from your business in each one. That average, multiplied by monthly volume, is your interaction baseline.
- Estimate how much of your prompt is fixed. Rules, catalog, and policies ride along on every call. A bloated prompt is a recurring cost disguised as configuration.
- Split what needs AI from what needs a template. Order confirmations, codes, and reminders don't need a language model: they're a template, they cost cents, and they don't get anything wrong. Save AI for the open conversation, where it actually changes the reply.
A good chunk of what today gets called "AI" is solved by deterministic automation. It's worth reading what separates one from the other in WhatsApp automation and in free WhatsApp chatbot, and checking the full API breakdown in how much the official WhatsApp API costs in 2026.
On CRPRO, AI is included in the monthly fee and runs on both the official-API number and the QR-code-connected one — if your operation is already past a few hundred conversations a month, it's worth comparing against your real number on the AI on WhatsApp page and on plans.
Frequently Asked Questions
How much does the Meta Business Agent cost per message?
About 4 to 5 US cents per interaction, somewhere between R$0.21 and R$0.26 at the August 2026 exchange rate. The stated price is US$2.00 per million tokens, and a message typically consumes 20,000 to 25,000 tokens, counting the prompt and the reply.
Does the per-token cost replace the message fee?
Yes, for messages answered by the agent. Meta says the token charge covers AI processing and delivery, and that a non-template message is charged either as a Meta Business Agent message or as a service message, never both.
Why does a short message consume 20,000 tokens?
Because what's charged isn't just the customer's text. It includes the agent's instructions, the catalog, the policies, and the conversation history — and the history gets resent on every new turn.
Is there a volume discount on the Meta Business Agent?
There's no announced volume discount for token consumption: the price per interaction is the same at any volume. A volume discount does exist on the message fee, for utility and authentication.
From how many conversations does AI inside the CRM become worth it?
By the math above, the break-even lands between roughly 2,100 and 2,700 monthly interactions — 300 to 390 seven-turn conversations. Below that, paying per use tends to be cheaper; above it, the gap grows in a straight line.
Does October 2026's charge change the token price?
What took effect on October 1 is per-message billing for service messages and utility templates inside the 24-hour window. The service rate is in Meta's current price table. The per-token price is a separate line, and it's been in effect since August 1, 2026.