Meta Business Agent Billing Is Live. Meta's Own Example Prices One Question at 20 Cents, Not 4.

August 3, 2026. Token billing for Meta Business Agent went live on Saturday. On Friday we published that the rate every outlet was quoting could not be found on a single Meta property, and we promised a Monday re-check. Here it is. The rate is now documented, it matches what was reported, and Meta's own worked example prices one simple customer question at about four times the number that made the headlines.

What Meta actually published

1. The rate is real, and it is global. Meta's developer documentation now states one global rate of $2.00 USD per 1 million tokens for Meta Business Agent messages, as a single charge covering both AI processing and message delivery. The page carrying it, Upcoming pricing updates for Meta Business Agent, service and utility messages, is stamped July 1, 2026 and is now linked from the top of the main WhatsApp pricing doc.
2. Meta puts one message at 20,000 to 25,000 tokens, which it converts to roughly 4 to 5 cents per message. That is the figure that circulated for three weeks before it appeared anywhere on a Meta property.
3. Meta's own example of a trivial question costs 16 to 20 cents. The doc prices the user message "At what time do you open?" at 4 messages to the user and about 80,000 tokens. Its complex example, a customer stuck on step 3 of an assembly, runs 10 messages and about 250,000 tokens, or 40 to 50 cents.
4. No credit line, no delivery. Meta states that as of August 1, Business Agent messages will not be delivered if a credit line is not attached to the Business Agent account. Credit card support is targeted for September, so the only route today is monthly invoicing set up through Billing Hub.
5. The free entry point window no longer covers everything. Replies inside the 72 hour window opened by a Click to WhatsApp ad remain free for message delivery, but Business Agent token charges apply inside it from August 1. In Meta's own comparison table, Meta Business Agent is the only message category marked "No" under free entry point treatment.
6. October 1 is the second and larger step. Service messages become chargeable per message for the first time since November 1, 2024, and utility messages sent in reply inside an open 24 hour customer service window become chargeable for the first time since July 1, 2025. Meta says it will publish those rates by September 1, 2026, and that service messages will carry no volume tiers.

The quoted price is per message, not per answer

This is the part the coverage missed, and it is not a technicality. Every summary of this change, including the one we published on Friday, described the cost as 4 to 5 cents per message. Meta's own examples are priced per interaction, and an interaction is not one message.

A customer asks what time you open. Meta bills that at four messages. Not because the agent is badly configured, but because that is how a conversational agent answers: an acknowledgement, the answer, a follow up offer, a close. At about 20,000 tokens each, the trivial question lands at 16 to 20 cents. The support scenario, ten messages deep, reaches 40 to 50 cents.

Meta is explicit that this is intentional. Its own note reads that not every Business Agent message costs the same, and that you pay proportionally to the value delivered. That is a defensible pricing philosophy. It is also a budgeting trap for anyone who multiplied their monthly inbound message count by five cents.

Do that arithmetic on your own volume before the first invoice rather than after it. Two thousand customer questions a month at Meta's simple example rate is $320 to $400, not $80 to $100.

The switch that turns off delivery

Below the rate card sits the line that will cause more damage this week than the price. From August 1, Business Agent messages are not delivered if no credit line is attached. Not queued, not downgraded to a plain reply. Meta's wording is that they will not be delivered.

For an agency running WhatsApp on behalf of clients, that is a silent failure mode with no obvious symptom on the client side. The customer sends a message, nothing comes back, and nobody sees an error. Card payment is not available until September at the earliest, so a business that only has a card on file has to set up monthly invoicing through Billing Hub, then confirm the Business Agent account is attached to that credit line and not merely to the ad account.

If you run this stack for clients, that is a check to run today on every account rather than a ticket for next sprint. Our GoHighLevel implementation work increasingly touches WhatsApp as an inbound channel, and this is exactly the kind of dependency that belongs in a written runbook instead of somebody's memory.

What it means for operators

1. Reprice on interactions, not messages. Pull last month's inbound conversation count, not message count, and multiply by 16 to 50 cents depending on how much your agent talks. That is your real August number.
2. Attach the credit line, then send a test message. Confirming the setting is not the same as confirming delivery. Send one real message from a phone that is not yours.
3. Shorten the agent, because length is now a line item. Every extra turn is a separate charge. An agent that answers in two messages instead of four halves the cost of the same conversation, and a long system prompt now carries a price on every single reply.
4. Set a spend alert before October. The October 1 change reintroduces charges on human replies and on utility messages sent in response to a customer, both free for over a year. Anyone who staffed a human handover partly because it was free is about to learn what it costs.
5. Recheck the ad maths. If Click to WhatsApp is a real acquisition channel for you, the reply is no longer free even inside the 72 hour window. Add the agent cost to your cost per lead before you scale spend.

What we got right and what we missed

Our Friday piece said the $2.00 figure could not be verified on any Meta property, and that anyone quoting it to a client was quoting an uncitable number. That was accurate on Friday. It is no longer accurate, and the correction is simply that the reported figure was right all along.

The wider lesson survives the correction, and it cuts both ways. Waiting for the primary source cost us three weeks of being early on a number that turned out to be correct. Publishing it as fact would have cost nothing in this instance and everything on the day a reported figure turns out to be wrong. What the wait actually bought was the detail: the four message example, the credit line requirement, the free entry point exclusion. None of that appeared in the secondary coverage, because none of it was reportable until Meta wrote it down.

If you are running customer facing automation on a platform that can reprice it underneath you, the durable habit is not scepticism about the headline number. It is reading the vendor's own worked example, because that is where the assumptions live. We build and run this kind of AI automation for businesses that would rather find the assumptions before the invoice finds them.