How much does an AI chatbot cost?
A chatbot has two prices: the build, which is scoped work, and the running cost, which is almost entirely model tokens plus messaging fees. The running cost surprises people by being small. Ten thousand support conversations on a cheap model come to roughly $37 in model usage at current list prices.
Prices checked:
The short answer
Stop budgeting for the model and start budgeting for the integrations. At list prices the model is often the cheapest line on the invoice; what costs real money is wiring the bot into your CRM, calendar and order data so it has something true to say.
| Cost line | How it is billed | Current figure |
|---|---|---|
| Model, cheap tier | Per million tokens in and out | Claude Haiku 4.5: $1 in / $5 out per million |
| Model, mid tier | Per million tokens in and out | Claude Sonnet 5: $2 in / $10 out per million |
| Model, top tier | Per million tokens in and out | Claude Opus 5: $5 in / $25 out per million |
| Worked example | 10,000 support conversations, ~3,700 tokens each | ~$37 total on Claude Haiku 4.5 |
| Repeated context | Cache read vs standard input | 0.1x the input price on a cache hit |
| WhatsApp replies | Per delivered message, by category | Free inside the 24-hour customer service window |
| WhatsApp outbound | Per delivered template message | Billed by recipient country and template category |
Sources: Claude API pricing (Anthropic docs) · WhatsApp Business Platform pricing · prices checked: Aug 26, 2026
What you actually pay for
A web chat assistant on your own site
No messaging fees at all, so the running cost is tokens and hosting. At a few thousand conversations a month on a cheap model this lands in the tens of dollars, and the build is where the budget actually goes.
The same assistant on WhatsApp
Replies inside the 24-hour service window are free, so an inbound-led support bot adds very little. Costs appear when you start the conversation yourself with a template, which is billed per delivered message.
Voice AI on a phone number
The most expensive channel, because speech recognition and speech synthesis bill per minute on top of the model. Worth it where the alternative is a missed call, which is exactly the case in the lead-response work we do.
The line people forget
A long system prompt and a big retrieved context ride along on every single message. Caching that repeated block drops it to a tenth of the input price, which matters far more at volume than switching model.
How we scope it
Almost every chatbot quote we are asked to sanity-check has the same shape: a large build number, a vague "plus API costs", and no volume assumptions. That is backwards. The API cost is the one line you can calculate exactly today, and Anthropic publishes a worked example for it: about 3,700 tokens per support conversation, so ten thousand conversations on Haiku 4.5 comes to roughly $37. You can do that arithmetic for your own volume in a minute.
What genuinely varies is everything around the model. Whether the bot can read an order. How many channels it answers on. Whether it has to hand off to a human mid-conversation and keep the context. Whether it works in one language or three. Those decide the build, and they also decide whether the thing is useful, because a chatbot with no access to your data is a FAQ page that types slowly. When we quote, the model line is usually the smallest number in the estimate, and we say so rather than hiding it in a bundle. See the systems this sits inside on the AI chatbots page, and what the messaging side costs in WhatsApp Business API pricing.