Meta's Business Agent moved to token billing on 1 August 2026 at $2 per million tokens — roughly 4–5 cents a conversation. Here's the maths at your volume.

What actually changed on 1 August?

Meta's Business Agent — the AI answering your WhatsApp, Instagram and Messenger messages — stopped being bundled and started being metered. From 1 August 2026 it bills on token consumption at $2 per million tokens, rather than per message. A typical customer conversation runs somewhere around 20,000 to 25,000 tokens, which puts each AI-handled chat at roughly 4 to 5 cents. The free window that opened on 1 July ran for thirty-one days and then closed.

That's the whole change. It sounds small, and per conversation it is. The reason it matters is that it removes the one thing flat pricing gave you for free: a bill you could predict without thinking about it.

What does that cost at your volume?

Do the arithmetic before you have an opinion about it. At ~22,500 tokens a conversation and $2 per million:

• 100 conversations a month → about $4.50. Rounding error.
• 1,000 conversations a month → about $45.
• 5,000 conversations a month → about $225.

Nothing there is frightening. But notice the shape of it: the cost scales with your success. A post that does unusually well, a seasonal rush, a link that gets shared into a group chat — each of those now has a small invoice attached. Under a flat fee, a busy month was free. Under a meter, a busy month is billed, and the bill arrives after the month rather than before it.

The honest read is that most solo businesses will pay less than they did on a per-message plan, and a minority with high inbound volume will pay noticeably more. Which one you are is a question of arithmetic, not opinion — and you can answer it in about ninety seconds with last month's conversation count.

It's also worth knowing what actually moves the token count, because it isn't the number of messages. It's length and complexity: how much of your product catalogue or FAQ the agent has to read to answer, how long the conversation runs, and how much back-and-forth it takes to resolve. A tightly scoped agent answering "what are your opening hours" burns a fraction of what a sprawling one burns working out whether a specific order qualifies for a specific refund. Which means the cost is partly a design decision, not just a volume one — and a well-briefed agent is now cheaper as well as better.

What's the second change nobody has read?

There's a second date in the same announcement, and it's the one worth diarising: from 1 October 2026, Meta resumes charging for service messages — the free-form replies businesses send inside the 24-hour customer service window. Those have been free since November 2024.

So the sequence for anyone running customer conversations on Meta's platforms is: free window opens 1 July, AI conversations become metered 1 August, human-sent service replies become chargeable again 1 October. Three pricing changes in one quarter, on a channel most small businesses treat as a fixed cost.

This is the actual news, and it isn't really about Meta. It's that the assumption underneath your messaging setup — replying to customers is free — has an expiry date on it now. Anything you built on that assumption is worth re-reading before October.

What should you do this week?

Three moves, in order, none of them dramatic.

Find your real number. Open your Meta Business Suite and count last month's AI-handled conversations. Multiply by $0.045. That's your baseline, and it's probably smaller than you feared — which is useful, because now you can stop worrying about it and start managing it.

Put a ceiling on it before you need one. Usage-based pricing has no brakes fitted as standard; the whole point of setting a hard spend cap in every metered tool is that it fails safe rather than failing expensive. Do it while the number is boring.

Change the question you ask. "What's the subscription?" has quietly stopped being useful. The number that survives metered pricing is cost per outcome — what one answered enquiry, one booked call, one resolved complaint actually costs you.

And keep the comparison honest. Four cents to answer a customer question at 11pm is cheap against any alternative that involves a person. The risk was never the unit price. It's that a meter runs whether or not anyone is reading the output — which is the same oversight problem that arrived when agents got access to your files, now pointed at your bank account instead of your documents.

Metered isn't worse. Unmeasured is worse. The tool that bills you per job is finally telling you what each job costs — which is information you never had on a flat plan, and information is the thing you can act on.

Not financial advice — figures are Meta's published rates as at 2 August 2026 and token use per conversation varies with length and complexity. Check your own dashboard before you budget.

For the vetted tool list we actually run — free, 5-min read: 10 AI Tools That Save 10 Hours a Week.

Related Decodes

• AI Just Became a Utility Bill — the week's hub: the full picture of metered AI and the 5-minute bill audit.
• Give Every AI Tool a Ceiling — the three-click fix, in every tool you pay for.
• The £0 AI Stack — the free-tier version of this question, and the price tags hiding inside it.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

TL;DR: Meta's Business Agent moved from bundled to token billing on 1 August 2026, charging $2 per million tokens — about 4–5 cents per AI-handled conversation across WhatsApp, Instagram and Messenger, since a typical chat uses 20,000–25,000 tokens. At 1,000 conversations a month that's roughly $45. A second change lands on 1 October 2026, when Meta resumes charging for service messages that have been free since November 2024. The practical response is to count last month's conversations, set a hard spend cap now while the number is still small, and switch from tracking cost per seat to cost per outcome.