Meta Business Agent vs Your Own AI on WhatsApp: What Each Costs Per Conversation

Short answer: Per token, Meta’s WhatsApp Business Agent and a well-built custom agent cost roughly the same. Meta charges $2 per million tokens; a capable mid-tier model from Google lists at a comparable blended rate. The bill diverges because of how many tokens each reply consumes. Meta says a typical agent message uses 20,000 to 25,000 tokens, about 4 to 5 US cents, and you cannot tune that. A custom agent that reads only what it needs can answer for a fraction of that, but you pay to build and run it. Volume decides which wins.

How does Meta charge for its WhatsApp Business Agent?

By the token, at a single global rate of $2 per million tokens, counting both the customer message the agent reads and the reply it writes, with AI processing and message delivery bundled into one charge. Meta’s own documentation puts a typical message at 20,000 to 25,000 tokens, or roughly 4 to 5 US cents, and confirms that these messages are not also billed as service messages (Meta for Developers). Token billing for the agent started on 1 August 2026.

Business Today converted the rate at about ₹189 per million tokens (Business Today), which puts a typical agent message at roughly ₹3.80 to ₹4.70. We use that conversion throughout.

What does a custom WhatsApp AI agent pay per message?

Two separate charges: the model provider for the tokens, and Meta for delivering the reply as a service message. Delivery is free for the first 1,000 service messages per phone number each month, then billed per message at the India service rate (YCloud). The model cost depends on which model you pick and how much each reply reads, which is the lever Meta’s agent does not give you.

For a worked example we priced Google’s Gemini 3.8 Flash at the list price that applies from 1 January 2027, $1.50 per million input tokens and $7.50 per million output tokens (Google AI for Developers), deliberately using the higher post-introductory rate. A reply that reads 4,000 tokens of conversation and retrieved knowledge and writes 200 tokens costs about $0.0075 in model fees, roughly ₹0.71, plus ₹0.115 for delivery once the free 1,000 are used.

So which is cheaper per conversation?

At the same token count they are close; at typical token counts the custom agent is several times cheaper per message. The gap is almost entirely context: how much of the conversation, catalogue, and instructions each reply re-reads. That is a design choice in a custom agent and a black box in Meta’s.

Our example reply uses 4,200 tokens and costs about ₹0.82 including delivery. Meta’s agent at the 20,000-token low end of its own typical range costs about ₹3.80, roughly four and a half times as much.

Here is the monthly picture, assuming five AI replies per conversation, before GST, and excluding the cost of building and running a custom agent:

Conversations / month	AI replies	Meta Business Agent (₹3.80–₹4.70 each)	Custom agent, model + delivery (illustrative)
300	1,500	₹5,700–₹7,100	~₹1,120
1,000	5,000	₹18,900–₹23,600	~₹4,000
3,000	15,000	₹56,700–₹70,900	~₹12,200

The custom column is an illustration of one well-scoped design, not a quote. A sloppier agent that stuffs the whole catalogue into every reply can cost as much as Meta’s.

What does the per-message comparison leave out?

The fixed costs, and they decide the answer at low volume. A custom agent has to be designed, connected to your systems, hosted, monitored, and maintained, and that costs money every month whether you send 500 messages or 50,000. Meta’s agent has none of that overhead: you switch it on and pay per message. The useful way to read the table is that the difference between the two columns is the monthly budget a custom agent has to fit inside to be worth it.

When is Meta’s own agent the right choice?

When your volume is modest and your questions are simple enough to answer from information Meta can hold. A business handling a few hundred conversations a month, mostly about hours, prices, and availability, will often find Meta’s agent cheaper all-in, because the per-message premium is smaller than any custom build’s running cost. We would tell a client that, and we cover the capability side of the decision in Meta’s WhatsApp Business AI agent: use it or build custom?

When does a custom agent pay for itself?

When volume is high, or when replies need your own systems. At a few thousand conversations a month the per-message gap in the table becomes large enough to fund a build and ongoing operation. Separately, any agent that needs to check live stock, look up an order, write to your CRM, or follow your escalation rules needs integration work that a hosted agent may not offer, and that need often decides the question before cost does.

How do I work out my own number?

Start from last month’s real conversation count and the average number of replies per conversation, then price both options. For Meta’s agent, multiply AI replies by 4 to 5 US cents. For a custom agent, estimate tokens per reply honestly, price them at your chosen model’s list rate, add delivery beyond the free 1,000, and add a realistic monthly figure for hosting and operation. If the gap does not comfortably cover that last line, Meta’s agent wins for now.

Our explainer on Meta’s WhatsApp AI pricing covers the billing mechanics in more detail. If you would like the comparison run on your own numbers, book a free 15-minute call, and if Meta’s agent is the better deal for you, that is what we will say. To see what a tuned custom agent sounds like, try our live WhatsApp AI demo.