How Much Does an AI Chatbot Cost? WhatsApp and Model Prices for Egypt and the Gulf

The monthly cost of an AI chatbot comes down to two meters you can read before you sign anything: how many messages pass through WhatsApp, and how many tokens the language model consumes. For an Egyptian store handling 1,500 conversations a month, the two meters together come to roughly $45 if the bot runs on a small model, and over $390 if it runs on the largest model available, with the same number of messages and conversations.

Before you calculate anything, note one date. On 1 October 2026 Meta starts charging for the service messages your bot sends for free today inside the 24-hour window. That single change raises the messaging bill of a typical support bot by about tenfold, as the worked example near the end of this article shows.

The message meter: how Meta charges today

Meta flipped its billing model on 1 July 2025: instead of pricing the conversation, it now charges for every template message delivered to the customer. Two things set the price: the template's category (marketing, utility or authentication), and the country calling code of the recipient's number, not your company's country. A store in Jeddah messaging a customer on an Egyptian number pays Egypt's rate.

Until 30 September 2026, two free doors stay open. First, your non-template messages inside the customer service window, the 24 hours that open and reset with every message the customer sends, are not charged, and neither are utility templates sent inside it. A bot that only answers people who wrote to you pays almost nothing today on the messaging line. Second, when a customer arrives from a "Click to WhatsApp" ad or a button on your page, a free entry point window opens and every message you send them is free for 72 hours, templates included.

What changes on 1 October 2026

Meta's documentation, updated on 10 September 2026, announces two changes that hit a customer support bot directly:

• Service messages become billable. Every non-template reply the bot sends inside the 24-hour window is charged at the same rate as utility and authentication templates in that market. Against that sits a new free tier of 1,000 delivered service messages per month for each business phone number, and whatever you do not use does not roll over to the next month.
• Utility templates inside the window become billable after being exempt since 1 July 2025.

For anyone planning beyond this quarter: Meta commits to changing prices only on the first day of a quarter (1 January, 1 April, 1 July and 1 October), with at least one month's notice for a rate card update, three months for an addition to the pricing model, and six months for a change to the model itself. That is the runway you have to adjust the bot's design before its bill goes up.

WhatsApp rate card by country

The figures below come from the rate card files Meta publishes on its pricing page, in US dollars, per delivered message (accessed 21 September 2026):

Kuwait and Oman sit inside the "Rest of Middle East" grouping today and move out to standalone rates on 1 October, along with Iraq and Morocco. Note that a single service message to a Kuwaiti number will cost more than twelve times what it costs to an Egyptian one, which matters if you serve several markets from one number.

Meta publishes the same card in sixteen currencies, including the Saudi riyal and the UAE dirham but not the Egyptian pound. In riyals, a marketing message to a Saudi number costs SAR 0.2159 from 1 October and a service message SAR 0.0401. In dirhams, a marketing message to an Emirati number costs AED 0.2115 and a service message AED 0.0576. An Egyptian account is billed in one of the published currencies rather than in pounds, so factor the exchange rate into your budget.

The token meter: language model prices

A language model is billed on input tokens (the customer's question, your instructions, the retrieved knowledge and the conversation history) and output tokens (the reply). The gap between the cheapest and the most expensive model from a single provider reaches fifty times on input.

Prices in US dollars, standard processing and short context, from the OpenAI, Anthropic and Google pages, accessed 21 September 2026.

Some of these prices are temporary by the provider's own wording: the Gemini pricing page states that input on Gemini 3.8 Flash is $0.75 until 31 December 2026 and $1.50 from 1 January 2027. If you build on a promotional price, put its expiry date in your calendar.

Alongside the replies, the bot consumes embedding tokens when it indexes your knowledge and on every search through it. Gemini Embedding 2 costs $0.20 per million text tokens on the same page, a small line that recurs with every catalogue update.

A worked example: a store with 1,500 conversations a month

What follows is a hypothetical example, not a price quote. The assumptions are written out in full so you can replace them with your own numbers:

• 1,500 conversations a month, all started by the customer, so a customer service window opens each time.
• The bot replies six times per conversation, which is 9,000 service messages.
• 900 utility templates for confirmation and shipping updates, sent outside the window.
• No marketing campaigns.
• Each model call consumes 3,000 input tokens and 200 output tokens.

The messaging bill for an Egyptian store, in dollars:

The same store in Saudi Arabia is billed in riyals: 8,000 service messages and 900 templates at SAR 0.0401 each, about SAR 357 a month from 1 October against SAR 36 today.

The model bill follows from the assumptions above: 9,000 calls at 3,000 input tokens give 27 million tokens in, and at 200 output tokens give 1.8 million tokens out per month.

Both meters together for this store from 1 October: about $45 a month on a small model and about $392 on the largest, before hosting, the vector database and the time someone spends reviewing conversations. If the quote in front of you carries a single monthly figure with no breakdown, these numbers tell you the size of what it leaves out.

Lines the two meters do not measure

• The one-time build cost, governed above all by integration depth. A bot that answers from static pages is far simpler than one that reads inventory, creates the order and issues a payment link through a custom system. Every integration adds development, testing and a new point of failure.
• The vector database. Qdrant Cloud offers a permanently free tier on a single node (0.5 vCPU, 1 GB RAM, 4 GB disk) that covers a small knowledge base, and a usage-based standard tier with no published price. Paid Pinecone plans start at $20 a month for Builder, with a $50 monthly usage minimum on Standard.
• Knowledge updates. Prices, policies and products change, and a bot answering from a stale catalogue costs you cancelled orders rather than invoices.
• Human review. Staff time to read a sample of conversations weekly and correct the weak answers.
• The provider's margin. Most companies here reach the platform through a business solution provider that adds its own fees on top of Meta's rate. Ask for the two prices separately in the quote.

How to cut the bill before the service meter starts

• Reduce the number of replies. While service messages were free, conversation length cost nothing; after the change, every extra message is a line on the invoice. A bot that collects the order details in one message with buttons is cheaper than one that asks a question per message.
• Count the free tier. A thousand service messages a month per business phone number. The tier is counted per number rather than per account, and a small store may never exceed it.
• Use the free entry point window. Sending customers to WhatsApp from a "Click to WhatsApp" ad makes 72 hours of messages free, templates included.
• A small model for simple tasks. Classification, pulling an order number and identifying intent do not need the largest model available; keep that one for the questions the small model fails.
• Cache your instructions. Anthropic's documentation states that a cache read is billed at 0.1x the input price, against a write at 1.25x for the five-minute cache. In our example, if two thirds of the input were fixed instructions and policies, the input line on Claude Haiku 4.5 would fall from $27 to about $11.
• Cut the excess context. Send only the knowledge relevant to the question and summarise long conversations. Tokens you do not send cost nothing.
• Ask about volume tiers. Meta offers lower rates on utility and authentication templates based on sending volume, and the tiers are published in separate files on the same pricing page.

What to ask for in the quote

Ask for four separate lines: the build cost per integration, the provider's monthly fees, an estimate of model usage in tokens rather than the phrase "depending on usage", and who pays Meta's invoice and owns the account. The last one becomes a problem the day you change vendors, because the WhatsApp Business account has to be in your company's name.

How to tell whether the bot earns its bill

Take from your WhatsApp history the number of repeat enquiries per month (price, appointment, order status), multiply it by the average handling time and the hourly staff cost, then add the enquiries that arrive after hours and go unanswered. Compare the result with both meters plus a share of the build cost. Start on one channel with a limited set of questions before expanding, so you measure your real numbers rather than assumptions.

To compare an off-the-shelf tool with a solution wired to your data before calculating any bill, see ChatGPT vs a custom AI solution. For the mechanics of building the bot itself on WhatsApp, its message templates and policy, see WhatsApp chatbot for your store.

Sources

• Meta: Pricing on the WhatsApp Business Platform, updated 10 September 2026, with the attached rate cards in dollars, riyals and dirhams.
• OpenAI: API pricing.
• Anthropic: Claude pricing.
• Google: Gemini Developer API pricing.
• Qdrant Cloud Pricing and Pinecone Pricing.

Frequently asked questions

What happens if I pass the free service message tier without a payment method on file?

Meta's documentation states that messages inside the free tier are delivered, while anything past it is not delivered at all as long as no payment method is registered in Billing Hub. In practice the bot can stop replying halfway through the month, so register a payment method before 1 October 2026.

How do I estimate token usage before launching?

Take a sample of twenty real WhatsApp conversations, run them through the model with the same instructions and retrieved knowledge you intend to use, then read the usage counter in the provider's console. Divide by the number of conversations for an average per conversation and multiply by your monthly volume. This is far closer than any published average, because the length of your instructions and the size of the retrieved knowledge are what drive the figure.

Does the bot handle Egyptian and Gulf dialects?

Modern models handle dialects, but the quality of the reply is set by what you test, not by what the vendor promises. Collect your customers' messages exactly as they wrote them, typos and abbreviations included, test the bot on them before launch, and add examples of the reply style you want to its instructions.

Do I need a large budget to try the idea?

No. The smallest useful experiment is one channel, the twenty questions you get most often, a small model, and no integration with inventory or the order system. That version reveals the share of questions the bot actually covers, and that share is what your expansion decision rests on.

We build AI and automation solutions at Snaabble from our office in Talkha, Dakahlia, for clients in Egypt and the Gulf. If you want the lines of a bot for your business broken out and its monthly usage estimated on your numbers rather than this article's assumptions, write to hello@snaabble.com or message us on WhatsApp at +20 103 673 3131.