AI agent cost has two parts that buyers often mix up: the one-time cost of building and connecting the agent, and the running costs of keeping it answering customers every month. The running costs themselves split into AI model usage, which is billed in tokens, and messaging fees, which depend on the channel, such as WhatsApp. This guide explains each part using the vendors' own public pricing pages, read on 12 November 2026, and shows a worked example with clearly labelled assumptions. It does not quote a build price for any provider, because build prices depend on scope and no source read for this guide publishes a fixed figure. CreativeIn Tek builds AI agents and does not publish a price list, so none is quoted here. All vendor prices change often, so treat every figure as indicative and check the page before budgeting.
Key takeaways
- AI agent cost has a build part (design, connection to business systems, testing) and a running part (model tokens, messaging fees, hosting and monitoring).
- Anthropic's pricing page lists prices per million tokens that differ by model; for example Claude Sonnet 5.5 is listed at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens.
- Meta's developer pricing page says that since 1 July 2025 it charges per message delivered from a business to a user, with rates that vary by message category and the recipient's country.
- The same page says that from 1 October 2026 each business phone number gets 1,000 free service messages a month, and that Meta resumed charging for service messages beyond that, but another Meta page says service messages are not charged, so check the current rate card.
- A model choice can change the token bill by more than twenty times, so test a cheaper model before defaulting to the most capable one.
The two halves of AI agent cost
The build half is mostly people's time. Someone has to decide what the agent should and should not do, write its instructions, connect it to the calendar, price list or booking system it needs, test it against real conversations and set up handover to a human. That work is billed by the project or by the hour, and it varies with how many systems are connected and how many rules the business has. The running half repeats every month and scales with the number of conversations. It includes the AI model, the messaging channel, the server that hosts the agent, and someone who reads conversations and improves the agent. Budgeting only for the build is the common mistake, because a busy agent can cost more to run in a year than it cost to build.
Token costs: what the vendor pages say
Large language models are billed in tokens, small pieces of text. Prompts and instructions sent to the model are input tokens, and the model's replies are output tokens. Anthropic's pricing page lists prices in US dollars per million tokens (MTok) and says all prices are in USD. The figures below were read on 12 November 2026 from the page's model table:
- Claude Sonnet 5.5: $2 per MTok input and $10 per MTok output.
- Claude Opus 5.5: $4 per MTok input and $20 per MTok output.
- Claude Haiku 5.5: $0.10 per MTok input and $0.50 per MTok output for prompts up to 100,000 tokens, and higher rates above that.
The page also lists a batch rate at half price for work that does not need an instant answer, and a lower price for cached input that repeats between requests. Other model providers publish their own pricing, and this guide did not read all of them, so a business should compare at least two providers for its own workload. The point for buyers is the spread: the cheapest and most capable models on one page differ by many times in price, and a customer-service agent that answers common questions may not need the most capable model.
A worked example with stated assumptions
Suppose, as an assumption and not a measurement, that one customer conversation sends 30,000 input tokens to the model in total, because the agent re-reads its instructions and the chat history with each reply, and receives 3,000 output tokens. Using the listed prices:
- Sonnet 5.5: 30,000 input tokens cost about $0.06 and 3,000 output tokens cost about $0.03, so roughly $0.09 per conversation, or about $90 for 1,000 conversations.
- Haiku 5.5: 30,000 input tokens cost about $0.003 and 3,000 output tokens cost about $0.0015, so roughly $0.0045 per conversation, or about $4.50 for 1,000 conversations.
These numbers are arithmetic on assumed token counts. Real conversations differ with message length, how much business data is placed in each request and how often the agent calls tools. A business should measure token use on a sample of real chats before trusting any estimate. Technical choices such as keeping instructions short and retrieving only the relevant business data per question usually reduce input tokens, and the guide to building a chatbot on business data explains the retrieval approach.
Scaling the example
Costs grow with volume, so it helps to see the same assumptions at a larger size. At 5,000 conversations a month, the Sonnet 5.5 example comes to roughly $450 and the Haiku 5.5 example to roughly $22.50, before messaging, hosting and staff time. If the agent handles simple questions with the cheaper model and passes harder ones to a more capable model, the bill can land between the two. That routing adds design work, so it pays off only when volume is high enough to justify it. A business with a few hundred conversations a month may find that people's time, not tokens, is the largest running cost.
WhatsApp and messaging fees
Messaging fees depend on how the agent connects to WhatsApp. Meta's developer pricing page says that, effective 1 July 2025, it charges on a per-message basis for the WhatsApp Business Platform, charges only for messages from businesses to users, and charges only when a message is delivered. Rates vary by message category, which the page lists as marketing, utility and authentication for templates, plus service for other messages, and by the recipient's country calling code. The page says it reflects updates effective 1 October 2026 and that rates are published in downloadable rate cards, which were not read for this guide, so no per-message rate is quoted here.
The page also describes changes to service messages. It says each business phone number gets 1,000 free service messages per month from 1 October 2026, with charges after that for most businesses. A separate Meta business page read the same day says that service messages are not charged, so the two pages conflict. The developer page is more specific and dated, but a business should confirm the current terms in its own Meta account before budgeting. The developer page also describes a Meta Business Agent product priced at a global rate per million tokens, which is a different product from building a custom agent.
A second route is a QR-linked connection to an existing WhatsApp account, which is how CreativeIn Tek's AI Agent connects. It does not use the official Business Platform, so it does not pass through Meta's per-message fees, but it also lacks the official platform's rules and protections. The trade-offs, including the risk of the number being restricted, are covered in the guide to the WhatsApp Business API versus a QR-linked agent. This guide does not claim the QR-linked route is approved by Meta.
Other running costs
- Hosting: the server or cloud service that runs the agent and stores conversations.
- Automation tools: workflow tools may add a subscription, for example n8n's pricing page lists plans from 20 euros a month billed annually.
- Connected systems: calendar, CRM or booking software subscriptions the agent depends on.
- Monitoring and improvement: someone has to read conversations, fix wrong answers and update information such as prices.
- Human handover: staff time for conversations the agent passes on.
A simple budgeting method
- Estimate volume: count customer conversations per month today.
- Measure tokens: run 50 to 100 real conversations through the agent in a test and record the tokens per conversation.
- Price two models: compare a capable model and a cheaper one on the same test and check answer quality.
- Add messaging: read the current rate card for the country and message type, or confirm what the connection method costs.
- Add hosting and tools, then add a monthly allowance for improvements.
- Set a ceiling: ask the builder to cap daily usage so a loop or abuse cannot create a surprise bill.
Questions to ask an agency before signing
- What is in the build price, and what counts as a change request?
- Who pays the model and messaging bills, the agency or the business, and are they passed through at cost?
- Who owns the agent, its instructions and the data, if the contract ends?
- How is quality measured, and how often are conversations reviewed?
- What happens when the agent is unsure, and how is a person brought in?
Businesses comparing agencies can use the same approach as in the guide to n8n automation agency costs in the UK, which shows how published prices were collected. CreativeIn Tek offers AI solutions as a service, and a business with a specific scope can ask for a quote.
Where an AI agent should not run alone
Cost is only one side of the decision. Agents should hand over to a person for complaints, payments, medical or legal advice, and anything irreversible. Marketing messages also fall under consent rules in each country, so a business should check those before automating outreach. A cheap agent that gives wrong answers costs more than a modest one that is reliable and escalates well.
Frequently asked questions
How much does an AI agent cost to run?
It depends on conversation volume, the model chosen and the messaging channel. A business can estimate it by measuring tokens on real conversations and adding messaging, hosting and maintenance.
How are AI models priced?
Providers bill per million input and output tokens, and prices differ widely by model. Anthropic lists its prices per million tokens on its pricing page, and other providers publish their own.
Does WhatsApp charge for AI agent messages?
On the official Business Platform, Meta charges per delivered business-to-user message at rates that vary by category and country. Its developer page says 1,000 service messages per number per month are free from 1 October 2026, so confirm current terms with Meta.
Is a QR-linked WhatsApp agent cheaper?
It avoids the platform's per-message fees, but it does not use the official platform and carries different risks, so the saving is not the whole picture.
Can the running cost be capped?
Yes. A builder can set usage limits, choose a cheaper model for simple questions and monitor spend daily.
Last reviewed: 12 November 2026, against the pricing pages of Anthropic, Meta WhatsApp Business Platform and n8n. Prices change often and are indicative only. Businesses planning an agent can talk to CreativeIn Tek.
