What a custom AI agent costs in 2026

A custom AI agent in 2026 costs one of two things: what an agency quotes, usually $3,000 to $15,000 for a single build or $4,000 to $10,000 a month for a retainer, or a fixed $9,500 to have one built, documented, and handed to you. Neither number tells you what it costs to run afterward, and almost nobody publishes that part. The real price has three lines: build labor, running cost, and the risk that the build gets scrapped before it pays for itself.

Search this exact phrase and you’ll find agency blog posts quoting $30,000 to $500,000 for a custom build. None of them link to an audited source, because none exists. What follows is built from public pricing pages, published API rates, and two 2025 studies on how often these projects actually get abandoned.

What does a custom AI agent actually cost in 2026?

Three line items make up the real number: build labor, running cost, and abandonment risk. Agencies quote $3,000 to $15,000 for a single build and $4,000 to $10,000 a month for a retainer. A full-time AI/ML hire runs $180,000 to $250,000 a year before ramp-up. We build one for $9,500 fixed, four to six weeks, and you own it at the end.

Path	Typical 2026 cost	Timeline	What you get
Agency, single build	$3,000 to $15,000	Weeks to months, hourly billed	One build, scope varies
Agency retainer	$4,000 to $10,000/mo	Ongoing	Ongoing service, no fixed end date
Full-time AI/ML hire	$180,000 to $250,000/yr + ramp	Months to hire, then ongoing	In-house capability, one person
AI workforce sprint (us)	From $9,500 fixed	4 to 6 weeks	One working tomte, code and accounts are yours

Prices as of July 2026. A tomte is our word for one production agent with one defined job. The agency and hire figures are the market comparison we publish on our own pricing page; the sprint price is ours.

The build labor line: agency, hire, or fixed price

Labor is the biggest line item and the one agencies are least specific about. Stack Overflow’s 2025 Developer Survey, with more than 49,000 respondents, put the US median AI/ML engineer salary at $189,500 and backend developer at $175,000. That’s salary alone, before benefits, before management time, before the months it takes to hire and ramp someone who can ship a working agent solo.

An agency avoids that fixed cost for you, but bills for its own version of it: $3,000 to $15,000 for a one-off build, or a retainer that runs $4,000 to $10,000 a month indefinitely. Neither model tells you, in writing, before you pay, exactly what you’ll own when it’s done. Ours does: $9,500 fixed, four to six weeks, scoped in writing before a dollar changes hands, and you get the repository, the database, and the hosting accounts, all in your name. Full detail is on the AI workforce sprint page.

The running cost line: what the token math actually shows

Inference is the cheap part. Anthropic’s published 2026 rates put Haiku 4.5 at $1 per million input tokens and $5 per million output tokens; Claude Opus 5 runs $5 and $25. OpenAI’s published 2026 pricing has gpt-5.6-luna at $0.20 and $1.20 per million; gpt-5.6-sol runs $5.00 and $30.00. Batch processing cuts the Anthropic numbers in half again.

Take a tomte that runs 2,000 times a month, a modest load for something like lead triage or reconciliation, at roughly 3,000 input tokens and 500 output tokens per run. That’s 6 million input tokens and 1 million output tokens a month. On Haiku 4.5, that’s $6 plus $5: about $11. On gpt-5.6-luna, about $2.40. Push it onto a heavier model, Opus 5 or gpt-5.6-sol, and the same volume costs $55 to $60 a month. Anthropic’s own worked example lands in the same range: roughly $37 to process 10,000 support conversations on Haiku 4.5.

Multiply any of those numbers by ten and you’re still under $600 a month. Running cost is rarely the reason a custom agent gets expensive, and it’s almost never the reason one fails. Labor sets the price. Abandonment is what erases the return on it.

The abandonment risk nobody prices in

Over 40% of agentic AI projects will be canceled by the end of 2027, Gartner said in June 2025, citing unclear business value and rising costs. Of the thousands of vendors marketing agentic capability, Gartner estimates only about 130 are genuine. Separately, S&P Global Market Intelligence found in 2025 that 42% of companies abandoned most of their AI initiatives that year, up from 17% the year before, and the average organization scrapped 46% of its AI proofs of concept before they reached production, in a survey of more than 1,000 respondents.

None of that shows up in an agency’s quote. A $9,500 build that gets canceled at month two costs you $9,500 plus every internal hour spent chasing a system nobody ended up using. Price the agent without pricing the odds it survives, and you’ve priced half the risk. We wrote about why most pilots die before they ship in Why your AI pilot died; scope kills more builds than technology does.

Do you need a true agent, or a cheaper fixed-sequence workflow?

Before you price a custom agent, check whether you need one. Menlo Ventures surveyed about 495 US enterprise AI decision-makers in November 2025 and found enterprise AI spend hit $37 billion in 2025, 3.2 times the $11.5 billion spent in 2024. Only 16% of what companies call agent deployments actually qualify as true agents: systems that plan, choose tools, and adapt mid-task. The other 84% are fixed-sequence workflows running under agent branding, a script that executes the same steps in the same order every time.

A fixed-sequence workflow is cheaper to build and cheaper to run, because there’s no reasoning loop spending tokens on every decision. If the job is “pull data from A, check it against B, write to C,” that’s automation, not an agent, and it should cost well under $9,500. If the job genuinely branches and judges case by case, that’s where the agent premium earns its keep.

Why nobody has published an audited cost band

No tier-1 publisher, not Gartner, not Forrester, not McKinsey, has published an audited custom-agent build cost band as of August 2026. Every $30,000-to-$500,000 figure floating around the search results is an agency quoting itself, with no methodology attached and nothing to check it against.

Ours is public because it isn’t hourly. It’s a fixed number we publish on our pricing page and hold to, the same discipline we apply to our own operations. When our ad platform reported a 4.87x return on spend and the ledger disagreed, we built a system to find the true number instead of trusting the dashboard. You can see exactly how that turned out on our ad-spend truth build: the real contribution-margin return was 0.33x to 0.50x, not 4.87x. That’s the same standard we apply to a price tag. Publish it, date it, let anyone check it.

What this means for your decision

Add it up and a $9,500 fixed build, running for $11 to $60 a month in tokens, is a fraction of a $189,500 median engineer’s annual salary and inside the range agencies already quote, minus the hourly uncertainty. The part that actually decides whether the money was well spent is the abandonment line: 42% of companies scrapped most of their AI work in 2025 alone. A fixed scope, written before you pay, is the cheapest insurance against becoming part of that number.

For the wider comparison across audits, content pipelines, and retainers, not just single builds, see AI automation agency pricing.

Describe the job you want automated at /start and you’ll get a written, fixed-price scope back within one business day. No discovery call, no meetings, just the number and the plan. Start async.