How Much Does It Cost to Run an AI Agent? Real Monthly Numbers for Small Businesses
Builders on this site report running agents for $18 to $200 a month. Here is where the money actually goes (models, hosting, voice minutes, orchestration), with a worked example you can adapt and prices as of November 2026.
What it costs to run an AI agent is mostly not the AI
A growth consultant on this site runs eleven Claude workflows across Gmail, Calendar, Notion and Slack and puts the whole running cost at 18 to 38 euros a month. A founder running seven agents on a single Mac mini says about $200 a month. A consultant with an email chief of staff reports $50 to $100 in API costs. Those are self-reported, and they are all far below what people expect when they hear "AI agent".
The surprise is where the cost sits. Language model tokens are often the smallest line. Voice minutes, the platform that runs your workflow, the tools you connect and above all your own time make up the rest. This guide breaks it down with current prices and a worked example, so you can estimate your own bill before you build.
Where the money goes
There are five buckets. The model itself, charged per million tokens. The orchestrator, meaning n8n, Make, a script on a server, or Claude's own scheduled tasks. Hosting, if you run it yourself. Telephony and voice, if it talks. And the tools it plugs into, such as a CRM or calendar, which you probably already pay for.
For models, Anthropic's pricing page lists these rates as of November 2026: Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, Sonnet 5.5 at $2 input and $10 output, and Opus 5.5 at $4 input and $20 output. Cached input is far cheaper, at $0.10 per million for Sonnet 5.5, and the Batch API takes 50 percent off if you can wait for results. A rough rule: a token is about three quarters of a word, so a million tokens is around 750,000 words.
For orchestration, n8n Cloud lists a Starter plan at 20 euros a month with 2,500 workflow executions and Pro at 50 euros for 10,000, on annual billing as of November 2026. Self-hosting n8n's community edition costs only the server. Hetzner lists cloud servers starting at 5.49 euros a month excluding VAT, though you should check the plan specs before running an agent on the smallest one. For voice, Retell AI lists pay-as-you-go from $0.07 to $0.31 a minute, with a default example of $0.11, and its own phone numbers at $2 a month.
Real numbers from builders on this site
The cheapest documented case is the consultant's eleven Claude workflows, at 18 to 38 euros a month. The same builder says it found 280 euros a month of forgotten subscriptions in the first week of expense tracking.
At the heavy end, the account in Running Five Healthcare and Software Ventures on Seven Agents for $200 a Month (builtwithagents.ai/strategy/seven-agents-200-month-clinic-saas-pharma) reports about $200 a month for seven named agents. The same account is honest about the costs that are not on the invoice: a 40-hour failure in a fallback system, eight zombie processes from a watchdog bug and an agent that marked unfinished work complete. The bill was $200. The babysitting was not free.
For a one-off build, a freelancer reports delivering a $6,400 client product in 11 days with under $40 of Claude Code spend (builtwithagents.ai/strategy/freelance-product-build-6400-claude-code-11-days). At the volume end, an accounting pipeline reports about 2,047 invoices a day for roughly $10 a day in model cost, which works out near half a cent per invoice. All figures are self-reported.
A worked example: an email and booking agent for a small clinic
Here is how you would budget a realistic agent. A physiotherapy practice gets 300 patient emails and messages a month. An agent reads each one, classifies it, drafts a reply from the practice's FAQ and booking rules, and puts it in a queue for the front desk.
Model cost first. Say each message uses 3,000 input tokens, including the instructions and FAQ, and 500 output tokens for the draft. Across 300 messages that is 900,000 input tokens and 150,000 output tokens. On Sonnet 5.5 at $2 and $10 per million, that is $1.80 plus $1.50, about $3.30 a month. With caching of the repeated FAQ it falls further. Even if you triple the usage for longer threads, you are near $10.
Orchestration: n8n Cloud Starter at 20 euros, or about 6 euros on a small server you maintain. The email and calendar tools they already pay for. Total running cost is roughly $25 to $35 a month, or about $10 if you self-host and do not count your hours.
Now compare a voice version. The same practice also wants calls answered, around 400 minutes a month. At Retell's example rate of $0.11 a minute that is $44, plus $2 for a number, so about $46 before any calendar or CRM tool. Voice is costlier per interaction because you pay for the audio, the speech models and the telephony together, not only the language model. If you want the numbers for phone answering specifically, see our AI receptionist cost guide.
Step by step: estimate your own bill
Step 1: Count the interactions
Write down how many messages, documents, calls or tasks the agent will handle in a month. Use last month's inbox or phone log, not a guess. If it is under 100 a month, the model cost will be a rounding error.
Step 2: Estimate tokens per interaction
Measure a typical case: instructions, the customer's text, any reference material you paste in, and the answer. Multiply words by 1.33 to get tokens. Reference material you resend every time is the usual culprit; cache it or trim it.
Step 3: Price the model with a small and a large option
Calculate the monthly total on a small model such as Haiku 5.5 and on Sonnet 5.5, using the per-million rates above. If the small model's drafts are good enough for classification and routing, use it there and reserve the bigger model for the reply that needs judgment.
Step 4: Add the orchestrator and hosting
Pick n8n Cloud, Make, or a self-hosted setup and add its monthly fee. Check execution limits: a workflow that triggers on every incoming email can burn 2,500 executions fast.
Step 5: Add voice, phone and tool fees
If the agent talks, multiply expected minutes by the platform's per-minute rate and add the phone number. Then list every tool that is not already in your stack and add its fee.
Step 6: Add a time line and a ceiling
Estimate two hours a month of maintenance as a minimum, more in the first quarter. Then set hard spending limits in your model provider account so a loop cannot run up a four-figure bill overnight.
What this does not include
Setup is the real cost for most people. The OpenClaw chief of staff on this site took about four hours to set up on a $600 Mac mini, with $50 to $100 a month in API costs afterward. If you pay an agency to build, one-off prices run into thousands of dollars; see our guide on what AI automation agencies charge.
Failures cost money too. A runaway loop, a retry storm or an agent that calls a paid API repeatedly can multiply your usage. Put a monthly cap on your model account, alert at 50 percent, and log every run.
When the cheap version is the wrong version
Do not pick the smallest model for work where a wrong answer costs more than the tokens saved, such as quoting a price, giving medical or legal information, or posting to a ledger. Do not self-host on a $6 server if nobody can fix it when it goes down on a Friday night. And do not judge an agent by its running cost alone: a $30 agent that saves the front desk five hours a month is worth keeping. A $5 agent nobody reads is not.
Frequently Asked Questions
How much does it cost to run an AI agent per month?▾
Builders documented on this site report anywhere from 18 to 38 euros a month for a set of Claude workflows to about $200 a month for seven always-on agents. A simple email agent handling 300 messages can cost around $25 to $35 including n8n and model usage, and a voice agent costs more because you pay per minute.
Is the AI model the biggest cost of running an agent?▾
Usually not. At Sonnet 5.5 rates of $2 per million input tokens and $10 per million output tokens as of November 2026, 300 messages cost a few dollars. Voice minutes, workflow platforms and your own maintenance time are typically larger.
How much does a voice AI agent cost per minute?▾
Retell AI lists $0.07 to $0.31 a minute depending on the voice and model you choose, with an example total of $0.11 a minute, as of November 2026. Add a phone number, which Retell lists at $2 a month, and any scheduling or CRM tool you connect.
Can I run an AI agent for free?▾
Partly. You can self-host n8n's community edition, and some model and voice platforms offer free credits to start, but servers, API usage and your time are not free. Plan for at least a small monthly bill once real traffic arrives.
How do I stop an AI agent from running up a huge bill?▾
Set a hard monthly spending limit in your model provider's console, alert at half of it, cap retries in your workflow, and log every run. Test with a handful of messages first, then watch the first week's usage closely.
Is it cheaper to build an agent or hire a virtual assistant?▾
For repetitive, rule-based work such as sorting email or drafting standard replies, the agent is usually cheaper to run, but it needs setup and supervision. For work that needs judgment and relationships, a person is still the better choice. Our comparison of AI agents and virtual assistants goes through the trade-offs.