What a WhatsApp AI agent costs in 2026

Two ways to run AI on WhatsApp, two price meters

Since 1 July 2026 Meta has had five message categories, and every message you send falls into exactly one of them. Two of those categories cover replies inside an open customer service window, and which one applies depends on what wrote the reply:

• Meta Business Agent. In Meta's words, "any non-template message powered by the Meta Business Agent Platform". This is Meta's own AI agent, and it has its own per-token price.
• Service. "Any non-template message not powered by the Meta Business Agent Platform. These can be powered by a 3rd-party AI solution or people." A reply written by an AI model you or your developer chose, such as Claude, is a service message, priced exactly like a reply typed by your team.

Messages customers send you are free under both. Meta "only [charges] for messages from businesses to users", and only when they are delivered.

Source: Meta, Pricing on the WhatsApp Business Platform.

What Meta Business Agent costs

Meta charges "one global rate of $2.00 USD per 1 million (M) tokens". The charge covers both the AI processing and the delivery of the message. Meta counts input and output tokens and does not charge for cache tokens.

Meta's own estimate is that "one message typically consumes 20,000–25,000 tokens, which translates to approximately 4–5 cents (USD) per message", and that the price "varies with complexity". Meta's examples:

Three things to know before you budget for it:

• No free messages. The 1,000 free service messages per number do not apply. Meta's page says "Free messages: Do not apply."
• No volume discount. "Volume tiers: Do not apply."
• Charged after ad clicks too. When someone messages you from a click-to-WhatsApp ad, most of your messages are free for a while, but "Meta continues to charge for Meta Business Agent messages in the FEP window."

Source: Meta, Charges for Meta Business Agent messages and Comparing Meta Business Agent to service messages. Checked 26 October 2026.

What a custom AI agent costs

A custom build has two running costs: what Meta charges for each reply, and what the AI model charges to write it.

1. Meta's charge: service messages

Since 1 October 2026 Meta gives "every business phone number" 1,000 free service messages per month. Unused ones do not roll over. From the 1,001st, each service message to a South African (+27) number costs $0.0095. Meta's volume tiers apply only to utility and authentication messages, so the service rate stays flat however many you send.

The free 1,000 are per number, and they cover every service reply, whether your team typed it or the AI wrote it. If your staff already send 600 replies a month, only 400 free ones are left for the AI.

Sources: Meta, what changed on 1 October 2026; South Africa rate card, linked from Rate cards and volume tiers.

2. The model's charge: tokens

Our platform uses Anthropic's Claude: Haiku 4.5 for quick sorting and short replies, Sonnet 4.6 where the reply needs more care. Anthropic's published prices, in USD per million tokens:

Source: Anthropic, Claude pricing. Checked 26 October 2026.

How many tokens a reply uses depends on how much you send the model: your instructions, the business facts it needs, the recent conversation, and any data pulled from your systems. For the examples below we assume 3,000 input tokens and 200 output tokens per reply. example On those numbers:

• Haiku 4.5: 3,000 × $1/M + 200 × $5/M = $0.004 per reply
• Sonnet 4.6: 3,000 × $3/M + 200 × $15/M = $0.012 per reply

If a reply takes two model calls, for example one to sort the message and one to draft the answer, add the cost of the second call. Long conversations and big prompts cost more per reply. Anthropic charges less for cached input (see the table), which can help where a long, fixed prompt is reused, but our examples do not count on it.

Worked examples

Each example counts AI replies sent on one South African number in a month. It assumes no other service messages use up the free 1,000, and uses the token assumptions above. example All figures are in USD, because Meta does not bill in rand.

The Meta Business Agent column uses Meta's own per-message estimate; your real token use may be higher or lower. The custom columns exclude Euphony's own fees: a R5,000 event storming session, the build (from R15,000 ex VAT, plus R7,500 per integration) and a monthly plan from R4,950 ex VAT, which includes a conversation allowance and AI usage. Voice-note transcription runs on a separate speech-to-text service and is not included.

What the examples show

• At low volume, the meters barely matter. At 500 replies the difference is under $25 a month, so a build fee will usually outweigh the saving. At that size, pick whichever option does the job.
• On short, focused replies, the custom route costs less per reply. After the free tier our example works out to about $0.0135 a reply on Haiku and $0.0215 on Sonnet, against $0.04–0.05 for Meta Business Agent. That depends on the token assumption. Meta's own comparison, for 10,000 "higher-complexity" AI replies in Brazil, estimates about 9–10 US cents a reply for a third-party AI sent as service messages, against about 4–5 cents for Meta Business Agent. Big prompts, long conversations and the largest models can flip the result, so cost your own case.
• Ad traffic tips it further. In a conversation that starts from a click-to-WhatsApp ad, Meta does not charge for service messages while the free entry point window is open (at least 72 hours; Meta's developer docs currently say up to 7 days). A custom agent then pays only the model. Meta Business Agent messages are still charged. See the free messaging window explained.

When each one fits

Meta Business Agent is likely the better fit when

• you send a few hundred AI replies a month or fewer, so the per-message price adds up to little;
• Meta's own agent does what you need without reaching into your own systems; check its current features on Meta's pages before you decide;
• you want a single Meta charge that covers both the AI and the message delivery, with no build fee. Meta says the cost per message "varies with complexity".

A custom build is likely the better fit when

• answers depend on your live data: order status from your ERP, open slots from your booking system, a quote from your pricing. We connect WhatsApp to those systems through their APIs (see integrations);
• the AI should hand over to people: Claude sorts the message and drafts the reply, and your team takes over in a shared inbox, at which point the automation pauses (see AI support with human handover);
• customers send voice notes: we transcribe them, with South African language support, so the team reads instead of listens (see voice-note transcription);
• you send thousands of replies a month, or a lot of your traffic comes from click-to-WhatsApp ads;
• the AI should do a job, not just chat: capture a lead into your CRM, take a booking, or send a Paystack or Yoco payment link.

One rule that limits what AI can do on WhatsApp

Since 15 January 2026, Meta only lets "AI Providers" offer general-purpose AI assistants on the WhatsApp Business Platform where Meta is legally required to allow it. South Africa is not on that list. So you can't run an open-ended ChatGPT-style assistant on a business number here. An agent that serves your own business, for support, bookings, orders or sales, is a different thing, and that's what both options on this page are for.

Source: Meta, AI Providers.

What Euphony charges

You connect your own WhatsApp account through Meta's Embedded Signup, and Meta bills you directly at Meta's rates, in US dollars or another supported currency, to the payment method on your account. We add nothing to Meta's message fees. Our platform meters Claude usage per client. We charge R5,000 ex VAT for an event storming session, which sets a fixed build fee, then from R15,000 ex VAT, plus R7,500 per integration, to build the agent, and from R4,950 ex VAT a month to run it. Monthly plans include AI usage. See pricing for Meta's full South African rate card.

Questions people ask