Meta Business Agent Pricing 2026: Per-Token Costs

Meta Business Agent is Meta's own AI agent, built directly into the WhatsApp Business Platform, and from August 1, 2026 it's no longer free. Meta charges a flat $2.00 USD per 1 million tokens, applied globally, and most single-message exchanges land somewhere around 4 to 5 cents once you factor in a typical token count.

That sounds simple until you try to budget for it. Token pricing behaves nothing like the per-message rates WhatsApp charges for everything else - marketing templates, utility templates, authentication messages. A quick FAQ reply and a ten-turn troubleshooting conversation cost wildly different amounts, and there's no flat per-message number you can multiply by monthly volume and call a forecast.

So this goes through how the token pricing works, Meta's own cost examples, and how the agent compares against running a third-party AI tool on top of WhatsApp, so you can work out which setup actually costs less for your conversation volume.

What Meta Business Agent is

It isn't a chatbot you plug into WhatsApp. It's Meta's own AI infrastructure sitting inside the WhatsApp Business Platform, able to handle a full customer conversation end to end - answering product questions, qualifying leads, and handing off to a human the moment a conversation gets too complex or sensitive for it.

It launched on July 1, 2026, open to any partner or directly integrated business through the API, and for that first month it was free. Charges begin August 1, 2026, and that's the date to circle if you're testing or building on it.

You're not locked into using it exclusively. It can run alongside a third-party AI tool or a human support team, handing conversations back and forth depending on what the moment needs. That's part of why the cost comparison against third-party tools is worth doing properly instead of assuming one is automatically cheaper.

The per-token model

Every other WhatsApp message type - marketing, utility, authentication, service - is billed per message. Meta Business Agent is billed per token, the same mechanism most large language model APIs use.

The rate is $2.00 USD per 1 million tokens, a single global rate not adjusted by country the way delivery rates are. A typical exchange, counting both what the user sends and what the agent generates back, consumes somewhere around 20,000 to 25,000 tokens. Run the math and that's roughly 4 to 5 cents per message on average.

Cost isn't fixed per conversation. It scales with how much the agent had to process and generate, so a short, simple answer burns fewer tokens than a long multi-turn explanation, and your bill depends on the shape of your typical conversation, not just how many you have.

One detail to understand before building anything: Meta applies a single charge covering both AI processing and message delivery. Run a separate third-party AI on top of WhatsApp instead and you typically pay two bills, the AI vendor's charge and Meta's delivery fee. That's a bookkeeping difference, not a cost difference. It doesn't change which setup is cheaper, only how many invoices you reconcile.

Meta's own cost examples

Meta published two example interactions with their actual token counts, and they show how wide the range gets.

A simple inquiry, someone asking "at what time do you open", typically resolves in four total messages and around 80,000 tokens, roughly 20,000 per message. That's approximately 16 to 20 cents for the full exchange.

A more complex one, someone stuck on step three of an assembly and needing a walkthrough, runs closer to ten messages and 250,000 tokens, about 25,000 per message. Around 40 to 50 cents.

The gap between those two, roughly 2.5x, is the whole point of a token model. You're not paying a flat toll per conversation, you're paying for how much thinking and generating the agent had to do. A business whose typical support conversation is quick & repetitive sits near the low end. A business whose typical conversation involves real troubleshooting or multi-step explanation trends toward the high end, consistently.

Against a third-party AI

Meta published a direct comparison based on 10,000 AI-powered messages sent to users in Brazil, and the numbers are worth understanding, since a flat per-token rate looks simple on paper but gives you less control than a setup built around your own traffic.

On a lighter, low-complexity setup, third-party AI comes out ahead clearly: around 2 to 3 cents per message, roughly $268 across 10,000 messages, once you add the AI vendor's charge to Meta's delivery fee. Meta Business Agent charges the same flat 4 to 5 cents per message no matter how simple the conversation is. There's no lighter tier to drop into. Every FAQ, every one-line question, every "what time do you open" is billed at the same rate as something far more complex.

That flat structure is the real limitation, and it's why a third-party setup gives you more control over total spend. Three things a well-built third-party AI layer can do that the flat rate can't:

• Route by complexity, so simple repetitive questions get handled cheaply and only the conversations that need heavier processing cost more
• Choose your own model, since you're not locked into one provider's pricing and different tasks can run on different models tuned for cost and accuracy
• Optimise as you scale, because as volume grows and you learn which queries repeat, the setup can be re-tuned to keep average cost per conversation trending down, while Meta's flat rate stays fixed however much pattern data you accumulate

Complexity affects the cost of any AI setup, not just Meta's. A third-party solution without complexity routing can also see per-message cost climb on harder conversations, which is exactly why routing matters. With it in place, most of your volume stays in the lower range, since the majority of real support traffic is the simple, repetitive kind to begin with.

If you want a cost structure built around your actual conversation patterns instead of one flat global rate, book a quick call with our team and we'll map it against your real traffic.

Token charges still apply inside the free window

This one catches businesses out. The 72-hour free entry point window - the free period after someone clicks a Click to WhatsApp ad or a Facebook call-to-action button - still applies for message delivery. It's unaffected by this update in that respect.

But if Meta Business Agent is the one responding inside that window, you're still charged for token usage from August 1, 2026. Delivery is free. The AI processing behind it isn't. That matters most for ad-driven lead qualification flows running through the agent, where it's easy to assume the whole window is cost-free when only the delivery half is.

Who this hits, and how

If you're running high-volume, multi-region support, test the agent right after billing starts on August 1, while volume is still manageable enough to isolate real per-conversation token costs. Don't extrapolate from Meta's Brazil example alone. Run a controlled pilot against a representative slice of your actual traffic, across your busiest markets, before committing broader volume.

If you're a smaller team, the real question is whether your typical conversation is simple enough that per-token pricing is overkill. If most of what customers ask is repetitive (hours, pricing, order status), a lighter automation layer that resolves the simple stuff before it needs agent-level reasoning is usually cheaper than routing everything through Meta Business Agent. Save the agent for the conversations that genuinely need multi-turn reasoning.

If you're the developer, analytics and webhook details for agent messages are still being finalised by Meta and are expected before charges begin on August 1, 2026. Build your cost-tracking dashboards around a dedicated category for these messages rather than folding them into existing reporting, and plan that separation now instead of retrofitting it in August.

Token pricing rewards businesses that understand their own conversation shape. A flat per-message estimate won't get you close to a forecast here, the way it does for templates. What will is pulling a sample of real customer conversations, estimating rough token counts from message length and turn count, and running that against the $2 per million rate before you commit volume. The businesses likely to be surprised by their first bill are the ones who assumed "AI agent" meant a predictable flat cost. It means paying for exactly how much the agent had to work, conversation by conversation.

Pull twenty real conversations, count the turns, and do the math before August. And pricing isn't the only thing changing on WhatsApp Business this year - usernames are too.

FAQ

How much does Meta Business Agent cost per message?

There's no fixed per-message rate. Meta charges $2.00 USD per 1 million tokens globally, and a typical exchange consumes 20,000 to 25,000 tokens, landing around 4 to 5 cents per message on average. Simple, short exchanges cost less. Complex, multi-turn conversations cost more, sometimes 2 to 3 times as much.

When does Meta start charging?

August 1, 2026. The platform launched July 1, 2026 and was free for that first month while businesses tested it through the API. Billing is invoiced monthly going forward, based on actual token consumption across all conversations the agent handles.

Is it cheaper than a third-party AI chatbot?

Not by default. Meta Business Agent charges the same flat rate whether a conversation is simple or complex, while a well-configured third-party setup routes simple, repetitive questions through cheaper processing and reserves heavier compute for the conversations that need it. Meta's own comparison shows a low-complexity third-party setup costing less overall. The advantage of a third-party layer is you can tune it to your traffic instead of paying one fixed rate for everything.

Do I still pay inside the free 72-hour Click to WhatsApp window?

Delivery stays free inside that window regardless of who or what responds, since the free entry point window covers delivery specifically. But if Meta Business Agent generates the response, you're charged for token usage from August 1, 2026, separately from delivery.

What's the minimum setup to estimate costs before committing?

Pull a sample of recent customer conversations and estimate rough token counts from message length and number of turns. Multiply against the $2 per million rate for a rough per-conversation cost, then compare to what you currently pay a third-party AI vendor plus Meta's delivery fee for the same volume.

Does this affect utility or marketing template pricing?

No. Template pricing - marketing, utility, authentication - continues under the existing per-message model, with rates set by market and category. Meta Business Agent pricing applies only to non-template, AI-generated replies handled through Meta's own agent, billed per token. The two systems run separately.