AI Agent Development Cost (2026): Build, Run and Maintain

A small, single-job AI agent costs roughly $500 to $5,000 to build with a freelancer or a fixed-price service, and from about $12,000 to well over $100,000 with a development agency. After that you pay to run it: API tokens, plus hosting, plus someone’s time when it breaks. For a small agent the token bill is often tens of dollars a month, not hundreds, but it scales with volume, so estimate it before you buy.

Every price below links to its source and was checked on October 25, 2026. Prices on vendor pages change often; treat them as a snapshot, not a quote.

Disclosure: Claudekit sells custom AI agents. Our prices appear as one row in the table below, read from our Hire page. We have tried to make this guide useful whether or not you buy from us, including a section on when you don’t need to pay anyone.

The three parts of the cost

Part	What you pay for	Who usually pays
Build	Scoping, code, connections to your tools, tests, handover	One-off, to whoever builds it
Run	Model tokens on each run, hosting, any paid APIs it calls	Monthly, to Anthropic and your host
Maintain	Fixing breakages when an API, a prompt or a model changes; adding features	Monthly retainer, or ad hoc

Most cost guides from agencies focus on the build. For a small business, the run and maintain parts decide whether the agent is still worth it after six months.

Build cost: public prices we found

These are published prices, not estimates. “Fixed” means the vendor states a price for a defined scope.

Who builds it	What you get	Published price	Source
Freelancer on Contra	Custom business workflow agent; custom MCP server; hourly AI agent work	$2,000; $1,500; $50/hour	Contra profile
Freelancer on Contra	One production Claude + n8n workflow, 1–2 weeks, one revision, 50% up front	$2,500 fixed	Contra service
Small studio (Sandlabs, Australia)	Agent or MVP build, 2–6 weeks, 30 days of bug fixes, you own the code	From A$5,000	Sandlabs pricing
Claude Code consultancy (eSEOspace)	Custom Claude Code project	From $3,000	eSEOspace
Development agency (Groovy Web)	Single-tool chatbot MVP; RAG agent; multi-agent system of 3–5 agents	$12K–22K; $28K–52K; $60K–120K	Groovy Web cost guide
Development agency (Groovy Web)	One MCP server; a multi-tool MCP server	$8K–18K; $15K–30K	Groovy Web MCP costs
Development agency (Space-O)	Simple rule-based agent; multi-agent system	$15K–40K; $150K–500K	Space-O cost guide
Claudekit (us)	A custom agent priced by feature count, built and tested by AI agents with no human review step	Starter (1–3 features) $790, Business (4–7) $1,900, Pro (8–12) $4,500; founding price for the first 10 clients $490 / $1,190 / $2,900; 13+ features quoted	Hire page

What the spread tells you:

• The gap is mostly scope and process, not the model. Agency ranges include discovery, project management, dashboards, compliance and support. A $2,000 freelancer job usually covers one workflow and a handover.
• The agencies say so themselves. Groovy Web’s MCP article notes that a connector to one well-documented API “is close to a weekend project for an experienced team”, and that the cost multiplies once you add authentication scoping and everything around the protocol.
• Fixed prices need a fixed scope. Every fixed offer above defines what one unit of work is: one workflow, one feature, one build. If you can’t write your job down that clearly, expect a discovery fee or an hourly quote.

How Claudekit’s price works

We include our own terms so you can compare them like any other row. Your number of features sets the tier, and a feature is one concrete job, such as “reads new orders from Shopify”, connection included. AI agents write the code and run automated tests for each feature; no person reviews the code before delivery. That is how the price stays fixed. You get a private GitHub repo you own and 2 revision requests after delivery. Starter is paid 100% up front; Business and Pro are 50% up front. Hosting, API costs and maintenance after the warranty (14 days on Starter, 30 on Business and Pro) are not included. Full terms, refunds and what we refuse are on the Hire page.

If a human code review matters to you, for example because the agent will write to systems that hold customer data, that is a reason to pick a vendor who includes one, or to review the repo yourself.

Run cost: estimate it before you buy

Running cost is the one part you can calculate in advance. For an agent that calls Claude through the API:

monthly token cost = runs per month × (input tokens × input price + output tokens × output price)

Per-token prices for the current models, from Anthropic’s pricing page (checked October 25, 2026, per million tokens):

Model	Input	Output
Claude Haiku 5.5 (prompts up to 100k tokens)	$0.10	$0.50
Claude Sonnet 5.5	$2	$10
Claude Opus 5.5	$4	$20

Worked example (our assumptions, not a measurement). An inbox agent handles 40 emails a working day, 22 days a month: 880 runs. Assume each run sends 4,000 input tokens (instructions plus the email) and gets 600 output tokens back.

Model	Per run	Per month (880 runs)
Haiku 5.5	4,000 × $0.10/M + 600 × $0.50/M = $0.0007	about $0.62
Sonnet 5.5	4,000 × $2/M + 600 × $10/M = $0.014	about $12.32
Opus 5.5	4,000 × $4/M + 600 × $20/M = $0.028	about $24.64

Real numbers will be higher than a single-call estimate for three reasons. Agents that use tools send several requests per task, each carrying the conversation so far. Thinking tokens are billed as output. And Anthropic reports that agents use about 4 times more tokens than chat, and multi-agent systems about 15 times more. Prompt caching and the Batch API pull the other way; our Claude API pricing guide explains both.

Plug your own numbers into the free API cost calculator, then set a monthly spend limit on the workspace in the Claude Console so a loop can’t run up a surprise bill.

For scale on the high end: Anthropic’s Claude Code cost docs report an average of about $13 per developer per active day across enterprise deployments, for a coding agent used all day. A narrow business agent that runs a few times an hour should cost far less.

Whose API key it runs on

Ask this before you sign. An agent you buy should call Claude with your API key, under your own account with Anthropic, so the usage is billed to you and you can see and cap it. Anthropic’s Claude Code legal page says developers building products should use API key authentication, and that companies running Claude Code in their products may not pay for, resell or intermediate Claude usage on their end users’ behalf. A builder who offers to run your agent on their own Claude subscription is offering something you can’t rely on.

Hosting is the other run cost. A scheduled agent can run on a small server or a serverless function you already pay for; check your provider’s current prices rather than trusting a number in a blog post, including this one.

Maintenance cost

Agents break when the things around them change: an API version, a login, a spreadsheet column, a model update. Published examples:

• eSEOspace lists maintenance plans from $250 a month (up to one hour of support) to $4,000 a month (up to 30 hours). Source
• Space-O says run and maintain costs commonly add $40,000 to $120,000 a year for a mid-complexity agent handling moderate volume. Source
• Claudekit includes a 14- or 30-day warranty and no maintenance after it. You own the repo, so you or anyone you hire can maintain it.

Budget for maintenance from day one. If nobody on your side can read the code, a monthly retainer is part of the real price.

When you don’t need to pay anyone

You can skip the build cost entirely in more cases than vendors suggest:

• The job is answering questions from your documents. A Claude Project with your files attached does this with no code. Projects work on every plan, including Free (up to five projects).
• The job is a fixed sequence with no judgment. “When a form is submitted, add a row and send an email” is an automation, not an agent. Anthropic’s own advice on building effective agents is to find the simplest solution possible, which “might mean not building agentic systems at all.”
• It’s one workflow and you can read code. Claude Code can help you write a small agent with the Agent SDK, and you can run it from cron with headless mode. Follow our step-by-step guide to building an AI agent with Claude, then get an API key.
• Your process isn’t stable yet. Automating a process that changes every month means paying to rebuild it every month. Run it by hand until the steps stop moving.

Pay someone when the agent touches several systems, when you need tests and a handover you can trust, or when your time is worth more than the build price.

How to compare quotes

Line up every quote on the same five questions:

1. What exactly is in scope? Ask for the feature or workflow list in writing.
2. Who owns the code, and where does it live at the end?
3. Whose API key and which account does it run on?
4. What does it cost per month to run, at your volume? Ask for the token estimate and the model.
5. What happens when it breaks in month three: warranty, retainer or hourly?

Our guide to hiring an AI agent developer has the full checklist, and AI agents for small business lists twelve common jobs with a rough feature count for each, so you can price your own list.