CyberFreezeDev- Home
- Blog
- How-to guides
- AI chatbot development cost: what drives the price
AI chatbot development cost: what drives the price
The cost of a custom AI chatbot comes from a short list of choices you can see in advance, plus running costs for model tokens and, on WhatsApp, Meta's message charges. Here is each one, with dated vendor rates.
The cost of AI chatbot development depends on a handful of choices you can see before anyone writes code: which channels it answers on, how many systems it connects to, how many documents it reads, which languages it speaks, how it hands over to a person, where the model runs, and how it is tested and maintained. Running costs come on top: model tokens for every reply and, on WhatsApp, Meta's per-message charges.
We do not publish prices, because two projects both called "a chatbot" can differ many times over in the work involved. This guide shows what moves the number, so you can compare quotes on the same terms.
What drives the build cost?
| Driver | Smaller build | Larger build |
|---|---|---|
| Channels | One website widget | Website, WhatsApp and Microsoft Teams, each with its own rules |
| Systems to integrate | None: answers from documents only | Reads and writes your CRM, order system or booking calendar |
| Documents | A few dozen clean PDFs | Thousands of files, scans that need OCR, per-team access rights |
| Languages | One | Several, including mixed messages such as Tamil and English |
| Human handover | A contact form | Live handover into your team's inbox, with the conversation attached |
| Hosting | A hosted model API | An open-weight model on your own GPU server |
| Evaluation | A short test set | A test set per language, rerun on every change |
Integrations usually move the number most. A RAG chatbot that only answers from your documents is a contained piece of work. One that checks an order status, books an appointment or updates a lead has to handle logins, errors, timeouts and the case where the other system says no. Each system adds design, testing and something else that can break.
Documents come next. Clean, current files are quick to index. Scans, duplicates and three versions of the same policy need cleaning first, and if different staff may see different documents, access control has to be built in before the model sees any text.
Languages and handover are smaller items that are easy to underestimate. Each language needs its own test cases, because a bot that answers well in English can still misread a mixed Tamil and English message. A good handover passes the whole conversation and the customer's details to a person, so the customer never has to repeat themselves.
Do you need a hosted API or your own GPU?
A hosted API (Claude, GPT, Gemini) costs nothing up front and bills per token. An open-weight model on your own server costs hardware, power and someone to run it, but heavy use does not grow the bill, and prompts stay inside your network. For most first chatbots the API is cheaper. Privacy rules or very high volume can change that. We compare the two in AI inference: API or your own server.
What does it cost to run? Model tokens
On a hosted API you pay for input tokens (your instructions, the retrieved passages and the question) and output tokens (the answer). On Anthropic's pricing page, checked 27 October 2026:
- Claude Sonnet 5.5: $2 per million input tokens, $10 per million output tokens.
- Claude Haiku 5.5 (prompts up to 100,000 tokens): $0.10 per million input, $0.50 per million output.
As an illustration, take a reply that sends 4,000 input tokens and returns 300. On Sonnet 5.5 that is $0.008 plus $0.003, about $0.011 per reply, or about $110 for 10,000 replies a month. On Haiku 5.5 it is about $0.00055 per reply, or $5.50 for the same 10,000. The real figure depends on your documents and questions, and Anthropic notes that token counts vary with the language and the model's tokenizer. Caching the fixed instructions cuts the input side further.
The model choice is made on your own test set, not on price alone: the cheaper model is the right one only if it passes.
What does WhatsApp charge?
Meta charges per message on the WhatsApp Business Platform, by message category and market. Messages your customers send you are free. When a customer writes, a 24-hour customer service window opens, and your bot's replies inside it are service messages. Since 1 October 2026 each business phone number gets 1,000 free service messages a month, and Meta charges from the 1,001st.
Meta's USD rate card effective 1 October 2026 lists these service rates per message:
| Market | Service message (USD) |
|---|---|
| United Kingdom | 0.022 |
| United Arab Emirates | 0.0157 |
| Singapore | 0.016 |
| Saudi Arabia | 0.0107 |
| North America | 0.0034 |
| India | 0.0014 |
A UK number sending 5,000 replies in a month pays for 4,000 of them: about $88. Each message counts, so a bot that splits one answer into three messages triples the bill. Marketing templates cost more (UK: 0.0635 per message). Our WhatsApp CRM and automation work is designed around these rules.
Evaluation and maintenance after launch
A chatbot is not finished at launch. Budget for:
- Re-indexing when documents change.
- Model changes. Providers retire older models, so the test set is rerun before you switch.
- Reviewing unanswered questions each month and adding the missing content.
- Monitoring handovers, errors and cost per conversation.
To check whether the chatbot pays for itself, put the hours it saves and the running costs into our AI automation ROI calculator.
Get a scope, not a guess
Tell us the channels, the systems, the documents and the languages, and we will come back with a written scope, a test plan and an estimate of the monthly running cost. Book a consultation and we will scope it.
