WhatsApp's other price change: AI replies are now billed by the token

We have written twice about the 1 October change, when free-form replies inside the 24-hour customer service window stop being free. That is the one everybody is planning for. There is an earlier change that got much less attention, and it is the more interesting of the two, because it changes the shape of the bill rather than its size.

Two meters, five weeks apart

From 1 August 2026, Meta Business Agent conversations are billed on tokens consumed rather than per conversation — reported at around two dollars per million tokens, roughly four to five cents per conversation, bundling the AI processing and the message delivery into a single rate. It applies across WhatsApp, Messenger and Instagram. Then from 1 October, service messages are charged per message again, at the per-message rate Meta already applies to utility and authentication templates, set by market.

Why token billing behaves differently

Per-conversation pricing is predictable: you can multiply it by your volume and put the number in a spreadsheet. Token pricing scales with how much the model reads and writes. A long system prompt, a generous slice of retrieved knowledge and a chatty answering style are now line items. Your bill becomes partly a function of your prompt design, which is a genuinely new thing for most teams to have to think about.

What to do before 1 October

• Measure how many service replies you actually send inside the 24-hour window today. Most teams have never counted, and it is usually more than they guess.
• Look at your prompt and your retrieval settings as cost inputs, not just quality inputs.
• Put a cost cap somewhere in the stack that stops spending rather than quietly absorbing it.
• Keep at least one channel you own — a widget or your own domain — where the only meter is the model's.

The structural point

None of this is Meta behaving badly; running a messaging network costs money. But it is a reminder of what renting a channel means. When the channel is somebody else's, the meter is somebody else's too, and it can change twice in one quarter. An agent that also answers on your own domain and inside your own site does not get repriced by a conference announcement.

The specifics above were announced at Meta's Conversations event in June 2026 and reported widely since; rates are set per market, so check the figures against your own rate card before you build a forecast on them.