How Much Does an AI Chatbot Cost to Build? Build, Run and Maintain Costs Explained
How to estimate an AI chatbot budget from official model prices and Canadian developer wages, with a worked monthly cost example and a scoping checklist.
Key takeaways
- Build cost is driven by scope, integrations and evaluation work, not by which model you pick.
- Mid-tier models from Anthropic and OpenAI list at US$2 per million input tokens and US$10 per million output tokens; small models list at US$0.10 and US$0.50.
- In our worked support-bot example, prompt caching alone cuts the mid-tier token bill from about US$600 to about US$300 a month at published cache prices.
- In Ontario, Job Bank lists a median wage of $48.08 an hour for software developers, which is the starting point for estimating build labour, before overhead.
Why AI chatbot price quotes vary so much
Search for chatbot pricing and you will find ranges from a few thousand dollars to several hundred thousand. Most of those ranges come from agencies with no published method, so they are hard to compare. The spread is real, though, because "chatbot" covers very different products: a widget that answers from a fixed FAQ, an assistant that searches hundreds of documents, and an agent that reads and updates records in your systems.
Chatbots are a mainstream use of AI. Statistics Canada reported in May 2026 that 19.2% of Canadian businesses used AI to produce goods or deliver services in the previous 12 months, and 28.2% of those used virtual agents or chatbots. Among all businesses, 10.6% named the cost of using AI as a barrier. The way to control that cost is to price the three parts separately instead of accepting a single number.
The three costs: build, run and maintain
Every chatbot budget breaks down the same way. Asking a vendor to quote each line separately makes proposals comparable and shows where the money actually goes.
- Build (one-time): discovery, conversation design, retrieval setup over your content, integrations, the chat interface, security review, and an evaluation set of real questions with expected answers.
- Run (monthly, usage-based): model tokens, tool fees such as web or file search, document storage, hosting, logging and monitoring.
- Maintain (ongoing): keeping source content current, re-running evaluations when the model or content changes, reviewing conversations, and handling escalations to people.
What drives the build cost
The build is mostly people's time, so the honest way to estimate it is hours multiplied by a rate. Job Bank, the Government of Canada's labour market site, lists hourly wages for software developers and programmers (NOC 21232) in Ontario of $30.29 low, $48.08 median and $77.40 high, based on 2023–2024 survey data updated in November 2025. At the median, one 40-hour developer week costs $1,923 in wages alone. Agency rates add benefits, overhead, project management and margin on top, so expect a quote well above the wage figure.
What multiplies the hours is scope. These are the factors we ask about first when scoping a chatbot, roughly in order of impact:
- Actions versus answers: a bot that only answers questions is far simpler than one that books, refunds or updates records, which needs permissions, confirmations and audit logs.
- Number of systems: each integration (CRM, help desk, e-commerce platform, single sign-on) adds design, testing and failure handling.
- Content condition: clean, current pages are cheap to index; scanned PDFs, conflicting policies and undated documents need cleanup first.
- Languages: an English–French or English–Japanese bot needs evaluation questions and content checks in each language.
- Risk level: anything touching health, money or legal commitments needs stricter testing, human handoff and privacy review.
How to estimate the monthly running cost
Model providers bill per million tokens, with separate prices for input (your instructions, retrieved documents and conversation history) and output (the reply). The monthly formula is: conversations × turns per conversation × (input tokens × input price + output tokens × output price) ÷ 1,000,000. Current list prices, in US dollars per million tokens: Anthropic's Claude Sonnet 5.5 is $2 input and $10 output, and Claude Haiku 5.5 is $0.10 and $0.50 for prompts up to 100,000 tokens. OpenAI's gpt-6.1-sol is $2 and $10, and gpt-6-luna is $0.10 and $0.50.
A worked example, using assumptions you should replace with your own measurements: 10,000 conversations a month, 4 turns each, 6,000 input tokens and 300 output tokens per turn.
- Mid-tier model at $2 / $10: $0.012 input + $0.003 output = $0.015 per turn, $0.06 per conversation, about US$600 a month.
- Same model with 4,000 of the 6,000 input tokens served from Anthropic's prompt cache at $0.10: about $0.0074 per turn, or roughly US$300 a month (cache writes add a little).
- Small model at $0.10 / $0.50: $0.00075 per turn, about US$30 a month.
Other running costs that are easy to miss
Tokens are rarely the whole bill. Both Anthropic and OpenAI charge US$10 per 1,000 web searches, plus the tokens of the search results. OpenAI's hosted file search costs US$2.50 per 1,000 calls and US$0.10 per GB per day of storage after the first free GB. If you run your own vector database, hosting replaces those fees. Add application hosting, logging with retention that matches your privacy policy, and error monitoring.
Token counts also depend on the model. Anthropic notes that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, and that token counts vary by language. Japanese and French text will not cost the same as English, so measure with the provider's token counting on real conversations before committing to a budget.
How to lower the cost without lowering quality
The cheapest chatbot is the one that sends each request to the smallest model that answers it correctly. That can only be proven with an evaluation set, which is why we build one before choosing a model.
- Put stable instructions and reference text first and cache them. Anthropic prices cache hits on Sonnet 5.5 at 5% of the base input price.
- Route simple questions to a small model and escalate only hard ones to a larger model.
- Retrieve fewer, better passages instead of pasting whole documents into every turn.
- Cap reply length and trim old conversation history, since output tokens cost five times input on these models.
- Move non-urgent work, such as summarizing transcripts overnight, to batch processing, which Anthropic discounts by 50%.
What to decide before you ask for a quote
A vendor can only give a reliable number when the scope is concrete. Preparing these answers usually shortens discovery and makes quotes comparable.
- The 20 to 50 questions the bot must answer, taken from real emails, tickets or chats, with the correct answers.
- Expected monthly conversations and languages.
- Which systems it reads from, and whether it may change anything in them.
- When and how it hands over to a person, and who owns the content it answers from.
- Privacy constraints: what personal information it may see, where data must be stored, and how long logs are kept.
- A monthly running-cost ceiling, so the design can be tested against it from the first prototype.
About the author
Jayson Hao
Founder of Innovation Trigger Lab and a University of Toronto Computer Science graduate with an AI/ML focus. He designs and ships production RAG systems, AI chatbots, web platforms and mobile products.
View profileFrequently asked questions
How much does an AI chatbot cost per month to run?
It depends on volume and model. At current list prices, 10,000 four-turn conversations cost about US$30 a month on a small model or about US$600 on a mid-tier model in our worked example, before search, hosting and staff time.
Is it cheaper to use ChatGPT or Claude for a chatbot?
At list price the comparable tiers are close: OpenAI's gpt-6.1-sol and Anthropic's Claude Sonnet 5.5 both list at US$2 input and US$10 output per million tokens. Real cost depends on token counts, caching and which model passes your evaluation set.
Can I build an AI chatbot without a developer?
A bot that answers from public FAQs can often be configured in an off-the-shelf tool. Once it needs your private data, integrations, actions or privacy controls, developer time is usually where most of the budget goes.
Why are prices in US dollars?
Anthropic and OpenAI publish API prices in US dollars, so Canadian and Japanese companies pay a currency-dependent amount. Developer wages from Job Bank are in Canadian dollars.
Sources and further reading
- 1.PricingAnthropic (Claude Platform Docs)
- 2.PricingOpenAI (API Docs)
- 3.Software Development Programmer in Ontario — WagesJob Bank, Government of Canada, November 19, 2025
- 4.Canadian Survey on Business Conditions, second quarter 2026Statistics Canada, The Daily, May 27, 2026
