Skip to content
Academy/AI News/Meta’s own AI now bills per token: what a Business Agent reply costs
AI NewsPrice change

Meta’s own AI now bills per token: what a Business Agent reply costs

Since 1 August Meta charges for Meta Business Agent — its AI that answers your customers — by the token. Here is what one reply costs, how to compare it with anything else, and what to ask before you switch it on.

AAssistantLabs Team
Academy 6 min read Updated Oct 2026

What happened

On 1 August 2026 Meta started charging for Meta Business Agent — its own AI, which answers your customers on your behalf. It is the first time Meta has priced a reply by how much the AI worked rather than by how many messages went out.

The price is $2.00 per million tokens, one global rate. Meta's own estimate is that a typical message uses 20,000–25,000 tokens, which comes to roughly 4–5 US cents.

This is not the 1 October change. That one — every ordinary reply becoming billable — is a separate story, covered in WhatsApp starts charging for replies. Business Agent has its own meter.

What is actually new

Three details, and the third one is the one that surprises people.

The meter is tokens, not messages

A token is a small piece of a word. Models are billed by how many they read and write. A short reply can be cheap and a long, complicated one expensive, so the price of a reply is no longer a fixed number. Meta charges only input and output tokens, not cached ones.

One charge covers the AI and the delivery

Normally you pay Meta for delivering a message and pay somebody else for the AI that wrote it. Here it is one line: Meta describes a single per-token charge that includes the AI's processing.

The free entry window does not make it free

When a customer messages you from a click-to-WhatsApp ad, Meta does not charge for your marketing, utility, authentication or service messages during the free entry point window. Business Agent messages remain charged inside it. If you run ads and switch Meta's AI on, the leads you paid to acquire are exactly the ones whose answers cost you.

What a reply costs

Tokens are the wrong unit for a decision. Convert them into the thing you already count: the price of one reply.

At ₪3.10 to the dollar, 4–5 cents is about 12–16 agorot per Business Agent message. For comparison, an ordinary reply from your own team or another agent costs, in Israel, $0.0053 — about 1.6 agorot — and each business number gets its first 1,000 such replies a month free. So a Business Agent message costs roughly eight times an ordinary one, before you count what the AI behind the ordinary one costs.

A business with 300 conversations a month and six agent replies in each sends 1,800 replies:

  • Business Agent: about ₪225–280 a month, AI and delivery together.
  • Ordinary replies: the first 1,000 are free, the other 800 come to about ₪13 — plus whatever the AI that writes them costs.
That is not a verdict, because the two lines are not the same product. The second line leaves the AI's cost out. Ask what it is, per reply, before you compare. The price also says nothing about quality: an eight-times-dearer reply is worth it only if it answers better, and you will only know that by reading real conversations.

What stays the same

If you have not switched Meta's AI on, nothing changes for you today.

  • Your ordinary replies are untouched. Anything you or another agent sends that is not a template and is not from Business Agent is a service message, priced as above.
  • Nothing is forced on you. Meta's pricing page describes a category you can use; it does not describe one you are moved into.
  • Availability is unconfirmed. The pricing page gives a single global rate and does not say in which countries Business Agent is offered. Check inside your own account before you plan around it.

What to do this month

An hour, and no spending. The aim is to be able to compare, not to decide.

  • Find out whether it is on for you. Look in your WhatsApp Business account for Meta's AI. If it is not there, you can stop here.
  • Ask every alternative the same question. “What does one ordinary customer reply cost, with the AI included?” A per-seat or per-token answer is not an answer.
  • Price a real conversation, not an average. Open one from last week, count the replies, multiply by the figure above.
  • If you try it, test it first. Run it on a few dozen real conversations and read what it said. Look for invented prices, promised callbacks and answers about things you do not sell.
  • Keep your ad leads out of it, for now. Those are the conversations where a metered AI costs money that the rest of your WhatsApp does not.
One thing not to do: do not choose by price alone. The cheapest reply that sends a customer away costs more than the dearest one that closes.

How to read the next one

Usage-based AI pricing will keep arriving, from Meta and from everyone else. A few questions make any of it readable.

  • What is the unit? Tokens, messages, seats and conversations are not comparable until someone converts them. Do it yourself, into a price per reply.
  • Who is the typical customer in their example? A “typical” message in a vendor's estimate is usually a short one. Yours may not be.
  • What is bundled and what is not? Here delivery and AI are one charge. Elsewhere they are two.
  • Is the date a test or a bill? This one had a free window through July. A start date is when the meter runs.

The pattern to carry forward: AI replies are becoming a metered utility. The skill is not finding the cheapest meter. It is knowing your own number, per conversation, so that you can tell when a price is fair.

Was this helpful?