AI agent cost calculator:
build plus running costs.
Enter your build quote, your monthly conversation volume and the channels the agent uses. The calculator adds model token costs, Meta’s WhatsApp fees for Ireland, Twilio phone costs, hosting and maintenance, and shows the monthly running cost and the first-year total in euro.
What makes up the cost of an AI agent?
An AI agent has a one-time build cost and a set of running costs that grow with use. The build cost comes from your supplier’s quote. The running costs come from the services the agent uses every time it works.
| Item | How it is charged | Where the price comes from |
|---|---|---|
| Build | One time | Your quote. We do not publish ranges because no public source gives reliable ones. |
| Language model | Per million input and output tokens | Each vendor’s pricing page, in US dollars |
| Per delivered message, by category | Meta’s rate card for the market that includes Ireland, in euro | |
| Phone | Per number per month, per minute | Twilio’s Ireland voice pricing, in US dollars |
| Hosting | Per month | Your hosting plan or quote |
| Maintenance | Per month | Your support retainer, or the staff time you set aside |
The AI agent cost guide explains what a build quote should include and why two quotes for the "same" agent can differ.
Estimate your agent’s costs
Change any value. The totals update as you type.
Build once, then run every month.
The starting values are an example, not a benchmark or a neu figure. Prices come from the vendors’ own pages (see Sources). Model prices are in US dollars and we convert them at the ECB reference rate of 19 November 2026.
Build
Model use
WhatsApp and phone
Hosting and support
- Model (Anthropic Claude Sonnet 5.5)US$2 in, US$10 out per million tokens
- €18
- WhatsApp message feesMeta’s Ireland rates; incoming messages are free
- €45
- Phone line and minutes
- €0.00
- Hosting
- €0.00
- Maintenance
- €0.00
- Total per month
- €64
Estimate only, before VAT. It assumes one WhatsApp number. It leaves out any speech or voice fees that ConversationRelay does not cover, Twilio’s own WhatsApp fee (US$0.005 a message) if you send through Twilio, prompt caching and batch discounts, and currency movement. The AI agent cost guide explains what a build quote should include.
How the calculator works
All results are estimates before VAT. The calculator uses these formulas:
- Model cost per month = conversations × (input tokens × input price + output tokens × output price) ÷ 1,000,000, then converted from US dollars to euro at the ECB reference rate of 19 November 2026 (US$1.0992 to €1).
- WhatsApp per month = replies above the first 1,000 × €0.0142, plus marketing templates × €0.049, plus utility templates × €0.0142.
- Phone per month = numbers × US$1.80, plus minutes × (US$0.01 inbound + US$0.07 for Twilio ConversationRelay if you choose it), converted to euro.
- Running cost per month = model + WhatsApp + phone + hosting + maintenance.
- First year = build quote + 12 × running cost per month.
Tokens are the units models charge for. About 750 English words is roughly 1,000 tokens. The input count for one conversation includes the agent’s instructions, the conversation so far, any documents it reads and the results of the tools it calls, on every turn. That is why the default input is much larger than the output.
A worked example: 1,000 WhatsApp conversations a month
These are the calculator’s starting values. They are an example, not a benchmark. A business has 1,000 WhatsApp conversations a month. Each uses about 6,000 input tokens and 800 output tokens. The agent sends 4,000 replies and 200 utility templates (for example, booking confirmations) each month.
| Item | Calculation | Per month |
|---|---|---|
| Claude Sonnet 5.5 | 1,000 × (6,000 × US$2 + 800 × US$10) ÷ 1M = US$20 | €18.20 |
| Claude Haiku 5.5 instead | 1,000 × (6,000 × US$0.10 + 800 × US$0.50) ÷ 1M = US$1 | €0.91 |
| WhatsApp replies | (4,000 − 1,000 free) × €0.0142 | €42.60 |
| WhatsApp utility templates | 200 × €0.0142 | €2.84 |
Two points stand out. First, at this volume WhatsApp fees cost more than the model. Since 1 October 2026, Meta charges for business replies after the first 1,000 each month, so a busy agent pays for its replies. Second, the choice of model changes the model cost by a factor of 20, but the total changes much less. The larger costs are usually maintenance and the build.
For a phone agent, add minutes. 600 calls of 3 minutes is 1,800 minutes. With Twilio inbound and ConversationRelay, that is 1,800 × US$0.08 + US$1.80 for the number = US$145.80, or about €132.64 a month.
What the estimate leaves out
- VAT. Most vendor prices are shown without VAT.
- Discounts. Prompt caching cuts the price of repeated input, often by 90% or more. Batch processing is about half price for work that can wait. Real model bills are often lower than the list price.
- Twilio’s WhatsApp fee. If you send WhatsApp messages through Twilio, Twilio adds US$0.005 per message, inbound or outbound, on top of Meta’s fee.
- Speech and voice. Some voice setups use separate speech-to-text and text-to-speech services. Add their cost to maintenance or ask your supplier.
- Your team’s time. Reviewing handoffs, updating answers and checking logs takes staff time each week.
- Price changes. Meta may change WhatsApp rates on the first day of each quarter. Google has announced that Gemini 3.8 Flash prices double on 1 January 2027.
For the WhatsApp side in detail, read WhatsApp Business pricing in Ireland. To compare models side by side, use the AI model cost calculator.
Questions to ask your supplier about running costs
A build quote often says little about the monthly bill. Ask these questions before you sign, and put the answers into the calculator:
- Which model or models will the agent use, and at which step?
- How many input and output tokens does one typical conversation use? Ask for numbers from a test run, not a guess.
- Whose account pays the model, WhatsApp and phone bills: yours or the supplier’s? If the supplier pays, what mark-up do they add?
- How many WhatsApp messages will the agent send per conversation, and which are templates?
- What does the maintenance fee cover: model updates, answer changes, monitoring, bug fixes, or only hosting?
- What happens to the cost if volume doubles?
- Who owns the code, the prompts and the data, and can you move them to another supplier?
We recommend that the accounts for the model, WhatsApp and phone are in your business’s name. You then see the real usage, the bills go to you at cost, and you keep control if you change supplier. The build vs buy guide covers ownership in more detail.
Also ask how the agent behaves when a service fails. If the model provider is slow or down, the agent should tell the customer and pass the request to a person. That design choice costs little at build time and saves lost enquiries later.
How to keep the running cost down
- Use a small model for simple steps (routing, extracting details) and a larger one only where the task needs it.
- Keep the agent’s instructions short and load documents only when the question needs them.
- Use prompt caching for the fixed part of every request.
- Send fewer WhatsApp messages: one complete reply costs less than three short ones.
- Hand over to a person early on phone calls that go off script, because minutes cost money.
Neu builds AI agents and WhatsApp agents with these limits in the design. Our own FreeLLMAPI project puts many model providers behind one key.
Common questions
Why does the calculator not show build price ranges?
We found no public, primary source for AI agent build prices that we could cite. Prices depend on scope, integrations and testing, so a range without a source would mislead. Type the quote you have. If you have two quotes, run the calculator twice and compare the first-year totals.
Are WhatsApp replies free inside the 24-hour window?
Not any more. Since 1 October 2026, each business phone number gets 1,000 free service messages a month. After that, Meta charges for each delivered service message: €0.0142 in the market that includes Ireland. Utility templates sent inside an open window are also charged since that date.
How accurate are the token numbers?
They are a starting point. The real count depends on the length of the agent’s instructions, the conversation and the documents it reads. Ask your supplier for token counts from a test run, or check the usage page of your model provider after a week of use, then put those numbers in.
Why are model prices in US dollars?
OpenAI, Anthropic, Google, Mistral and most hosts list API prices in US dollars. The calculator converts them at the ECB reference rate of 19 November 2026. Your bill depends on the rate on the day you pay and on your card or bank charges.
Sources
- OpenAI: API pricing
- Anthropic: Claude API pricing
- Google: Gemini API pricing
- Mistral: API pricing (docs)
- Together AI: pricing
- DeepSeek: API pricing
- Meta: WhatsApp Business Platform pricing rules
- WhatsApp Business: platform pricing (rates by market)
- Twilio: Ireland voice pricing
- Twilio: WhatsApp pricing
- European Central Bank: euro reference rates
We last checked these sources on 19 November 2026. Rules, fees and prices change. Check the source before you act.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.