AI Agent Cost Calculator: know your real monthly bill before you commit

AI model prices compared

Model prices in our table run from $0.10 to $10.00 per million input tokens, and output costs more. A low-cost model suits simple jobs like sorting emails. Start with the cheapest model that might work, test it on real examples, and move up only when the answers fall short.

Here are the current API prices for the models in our calculator, per million tokens in US dollars. The last column shows what one support chat costs on each model, using our support chat preset (3 steps, 3,000 input and 400 output tokens per step, 50% cached). The table is sorted from cheapest to most expensive.

*Before any platform fee. DeepSeek prices are peak-hour prices; off-peak hours cost half. Google Gemini is not shown yet because we could not confirm its prices on the official page. Prices checked November 3, 2026.

Notice how prices cluster. GPT-6.1 Terra and Claude Sonnet 5.5 sit in the middle. GPT-6 Luna, Claude Haiku 5.5 and DeepSeek Flash (V4.1) are at the low end and are good for simple tasks like sorting emails. The top models, like GPT-6 Astra, cost ten times more and make sense for hard work where a mistake is expensive.

How to pick a model for your agent

Do not start with the most powerful model. Start with the cheapest model that might work, and move up only if you have to. Here is a simple way to test:

1. Collect 20 to 50 real examples of the task, like real customer questions or real leads.
2. Run them through a low-cost model and a mid-range model.
3. Score the answers yourself: right, almost right or wrong.
4. If the low-cost model gets nearly all of them right, use it. If not, use the mid-range model, or use the cheap one for easy cases and the bigger one for hard cases.

This one test often cuts the AI bill by half or more, because many tasks are easier than they look. Re-run the test when a new model comes out. New models are often both better and cheaper than the ones they replace.

Why the same model can cost different amounts

The list price is only part of the story. Two businesses using the same model can pay very different amounts because of prompt length, number of steps, how much of the prompt is cached and how long the answers are. Some providers also charge more for very long prompts, for data kept in a certain region, or for faster "priority" speed. Our calculator uses the standard price for normal-length prompts.

AI agent cost by tool type

AI agent tools charge in five main ways. Model APIs charge per token. No-code builders charge for runs or credits. Voice platforms charge per call minute. Chatbot builders charge per conversation or message. CRM agents charge a plan fee plus AI use. Match the billing to your job to find the best fit.

Each type of tool charges in a different way. Knowing how a tool bills helps you guess which one will be cheapest for your job.

4 real AI agent cost examples

In our examples, 100 support chats a day cost about $23.60 a month on the OpenAI API, and 40 voice calls a day cost about $217 on Vapi. Both use 22 working days and no safety buffer. Add 20% for a safe budget. Your numbers will change with your own usage.

These examples use the same math as the calculator, with 22 working days a month and no safety buffer, so you can follow every number. Add 20% on top for a safe budget.

Example 1: Customer support chat bot

A small online store gets 100 support chats a day. Each chat takes 3 model calls, with 3,000 input tokens and 400 output tokens per call. Half of the input is the same instructions every time, so it gets the cached price.

• Chats per month: 100 × 22 = 2,200
• Cost per chat on GPT-6.1 Terra: $0.0107
• Monthly cost with the OpenAI API: $23.60
• Same agent built in n8n Starter: $47.60 ($23.60 model + $24 plan)

Without caching, the same bot would cost $33.00 a month, or 40% more. Caching is one of the easiest ways to save.

Example 2: Voice receptionist

A dental office gets 40 calls a day. An AI receptionist on Vapi answers each one, and the average call lasts 3 minutes.

• Call minutes per month: 40 × 3 × 22 = 2,640
• Vapi hosting, speech, voice and phone: 2,640 × $0.0789 = $208
• AI model (GPT-6.1 Terra): $8.71
• Monthly cost: $217, or about $0.25 per call

Switching to GPT-6 Luna only drops the bill to $209, because the model was never the big cost. For voice agents, shorter calls save far more than a cheaper model.

Example 3: Lead follow-up in n8n

A real estate team gets 50 new leads a day. An n8n workflow reads each lead, writes a personal follow-up email with GPT-6.1 Terra and logs it in the CRM. Each lead is one workflow execution.

• Executions per month: 50 × 22 = 1,100, which fits n8n Starter (2,500 included)
• AI model: $6.03 · n8n Starter: $24
• Monthly cost: $30.03

At 150 leads a day you need 3,300 executions, which moves you to n8n Pro. The total becomes $78.08. Here the plan fee is most of the bill, so batching leads into fewer executions would save money.

Example 4: Research agent

A consultant runs 10 research reports a day. Each run takes 8 steps and reads 8,000 input tokens per step, because it reads web pages and notes.

• On Claude Sonnet 5.5: $35.06 a month ($0.16 per report)
• On Claude Opus 5.5: $70.12 a month
• On DeepSeek Flash (V4.1): $4.68 a month

Research agents use lots of input tokens, so input price matters more than output price here. Test a cheaper model on a few real reports before you commit to the expensive one.

What these examples teach

• Chat agents are cheap. The model cost for 100 chats a day is less than many software subscriptions. The plan fee of a no-code tool can be bigger than the AI bill.
• Voice is all about minutes. In Example 2, per-minute fees are 96% of the bill. A cheaper model barely moves the total.
• Plan limits create jumps. In Example 3, tripling leads more than doubled the cost because it crossed into a bigger n8n plan.
• Model choice matters most for heavy agents. In Example 4, the gap between DeepSeek Flash (V4.1) and Claude Opus 5.5 is $65.44 a month for the same work.

AI agent vs hiring staff in the US, UK, Australia and Canada

For simple, high-volume work, an AI agent usually costs far less than staff wages. In our example, 100 support chats a day take about 220 hours of a person's time each month. But an agent will not handle every case well, so keep people for the hard ones.

How does an agent compare to a person? Take Example 1: 100 support chats a day. If a person spends about 6 minutes per chat, that is 220 hours of work a month, or more than one full-time job. Here is what those hours cost in wages, next to the AI agent from Example 1.

US and Canada pay rates are official medians. UK and Australia are estimates we have not yet confirmed. Sources: US Bureau of Labor Statistics, Occupational Outlook Handbook, Office for National Statistics, Annual Survey of Hours and Earnings, Jobs and Skills Australia, occupation profiles, Government of Canada Job Bank, wage report. Wages only, not taxes or benefits, which add a lot to the real cost of a hire. Exchange rates from November 3, 2026.

The gap is huge, but it is not the whole story. An AI agent will not handle every chat well. Angry customers, refunds and odd questions still need a person. The smart way to use this number is not "replace the team," but "let the agent handle the easy 60 to 80% so your people can focus on the hard cases." Also count the setup time and someone to review the agent's work each week.

The real cost of a hire is higher than the wage

The table shows wages only. A real employee also costs payroll taxes, paid leave, training, equipment and management time. In many countries, that adds 20% to 40% on top of the wage. On the other side, an AI agent has costs the table does not show: building it, testing it, fixing it when a tool changes, and checking its answers.

Where agents beat people, and where they do not

Agents are best at work that is high volume, repetitive and has clear rules: answering the same 50 questions, copying data between systems, sending reminders and booking times. They work all night and on weekends, and they reply in seconds.

People are still better at anything that needs judgment, trust or a human touch: calming an upset customer, closing a big sale, handling a complaint or making a call when the rules do not fit. The best results usually come from a team where the agent does the first pass and a person handles the rest.

Why AI agent bills jump without warning

AI agent bills jump when an agent uses far more tokens than you planned. Each step sends the conversation so far, so long tasks get costly fast. A failed run that keeps retrying pays full price each time. Set limits on steps, retries and monthly spend before you go live.

Most people work out their budget from one model call. That is the usual mistake. Here is what makes the real number higher:

• Growing context. Each step is a new call that resends everything so far, so later steps cost more than early ones.
• Retries. A failed run that tries again pays full token price each time. Railway's cost guide gives an example: a 10% failure rate with 3 retries adds about 30% to the token bill.
• Prompts that grow. New tools, longer instructions and memory files make every call bigger over time. Measure again every few months.
• Launch numbers. Plan for your busy days, not your first week.

Hosting is rarely the problem. The same Railway guide says that for most agents the token bill is often about ten times the hosting bill. Its advice is the right habit for any agent: run it for a week, then replace your estimate with your real usage.

10 ways to cut your AI agent costs

The fastest ways to cut AI agent costs are a smaller model, prompt caching, fewer steps and shorter answers. For phone agents, shorter calls save the most. Set a monthly spending limit too, so a bug cannot run up a large bill. Try these one at a time and check the result.

1. Start with a smaller model. Try GPT-6 Luna, Claude Haiku 5.5 or DeepSeek Flash (V4.1) first. Move up only for the tasks where the small model fails.
2. Use two models. Let a cheap model sort and route requests, and send only the hard ones to a big model.
3. Turn on prompt caching. Keep your long instructions at the start of the prompt and the same every time. For most models in our table, cached input is 90% or more cheaper than normal input.
4. Cut the steps. Every step re-reads the context. If one well-written step can do the work of three, do it.
5. Trim the chat history. Send a short summary of old messages instead of the full history.
6. Ask for short answers. Output costs about 5 times more than input. Tell the model to be brief.
7. Use batch or off-peak pricing. If the work can wait, batch processing and off-peak hours can cut prices. DeepSeek charges half price off-peak, for example.
8. Keep voice calls short. For phone agents, minutes are the main cost. Get to the point, and hand off to text (SMS) for details.
9. Pick the right plan size. Check your real usage after a month. Many people pay for a bigger plan than they need, or pay overage on a plan that is too small.
10. Set spending limits and alerts. Every API provider lets you set a monthly budget. Use it so a bug cannot run up a huge bill.

Build vs buy: which is cheaper?

Building your own agent has the lowest running cost but needs a developer. A no-code tool costs more each month but is quicker to set up and change. A ready-made agent costs the most per task and is the fastest to launch. Many small teams should start with no-code or ready-made.

You have three main ways to get an agent running. Each costs money in a different place.

• Build with code and an API. Lowest running cost, because you only pay for tokens and a small server. But you need a developer to build it and keep it working. Best when you have high volume or very custom needs.
• Build with a no-code tool. Tools like n8n add a plan fee ($24 to $60 a month on the n8n Starter and Pro plans), but you can build and change workflows quickly without code. Best for most small and mid-size businesses.
• Buy a ready-made agent. Voice platforms and CRM agents charge per minute or per seat. They cost the most per task but are fastest to launch. Best when you need something working this week.

A simple rule: if the platform fee is more than half of your total cost, you may save by moving to the API directly once your volume grows. If developer time is your biggest cost, a no-code tool usually wins. At very high volume, check the bigger plans: n8n Business is $960 a month (self-hosted) for 40,000 executions.

What does it cost to build an AI agent?

Building is a separate cost from the monthly bill, and this calculator does not include it. With a no-code tool, you mostly pay with your own time. Agency prices vary a lot, so get more than one quote. As one public example, Musketeers Tech's build calculator starts at $6,000 for a single-task agent and $40,000 for a multi-agent system. Those are one agency's own prices, not a market average.

Questions to ask before you choose

1. How many tasks will the agent handle per day now, and in a year?
2. Does it need to talk on the phone, or is text enough?
3. Who will fix it when something breaks: you, a developer or the vendor?
4. Which tools does it need to connect to, like your CRM, calendar or inbox?
5. How bad is a wrong answer? If mistakes are costly, budget for a better model and a person to review.

Run your numbers through the calculator for each option. Then add the build cost and the cost of your own time. The cheapest monthly bill is not always the cheapest choice over a full year.

AI agent cost FAQ

How we get our prices

Every price comes from the tool's official pricing page, and we show the date we checked it. If we cannot confirm a price, we say so and do not guess. Usage numbers, like tokens per task, are our own averages. That is why your result is an estimate and not a quote.

Every price in this calculator comes from the tool's official pricing page. We record the page link and the date we checked it. If we cannot confirm a price on the official page, we mark it "Price not confirmed" instead of guessing. You can see the source for each tool on its own page.

Usage numbers for each task type, like tokens per step and minutes per call, are our own averages from typical agents. They are a starting point, not a promise. Exchange rates come from the European Central Bank reference rates on November 3, 2026.

Prices last updated: November 3, 2026. Official sources: OpenAI API, Anthropic Claude API, DeepSeek API, n8n, Vapi, xAI Grok API, Mistral API, GoHighLevel AI Employee, Retell AI, Bland AI, ElevenLabs Agents, Synthflow, Make, Zapier Agents, Lindy, Botpress, Chatbase, Voiceflow, Relevance AI.

Who this calculator is for, and its limits

We built this for small business owners, agencies and developers who need a real number before they commit. It suits you if you are comparing tools, pricing a client project or checking if an agent is worth building.

It has honest limits. We do not run your agent, so we cannot see your real prompts or call lengths. It leaves out build time, your own servers, paid tools like web search, and the person who checks the agent's work. Some links are affiliate links, including GoHighLevel. They never change the numbers.

Get your AI agent cost estimate

You now know what drives the bill, so put in your own numbers. The AI agent cost calculator gives you a monthly figure in about a minute. Looking at one product? Find its calculator here. Then run the agent for a week and compare your real usage with the estimate.

More calculators: OpenAI cost calculator, Claude cost calculator, DeepSeek cost calculator, Gemini cost calculator, n8n cost calculator, Vapi cost calculator, Grok cost calculator, Mistral cost calculator, GoHighLevel cost calculator, Retell AI cost calculator, Bland AI cost calculator, ElevenLabs cost calculator, Synthflow cost calculator, Make cost calculator, Zapier Agents cost calculator, Lindy cost calculator, Botpress cost calculator, Chatbase cost calculator, Voiceflow cost calculator, Relevance AI cost calculator.