How much does it cost to build an AI product in 2026

Ask three different shops to quote the same AI product and you will get $40,000, $400,000, and “it depends.” Nobody is necessarily lying. They’re pricing different things. Until you can see the line items underneath a quote, you can’t tell an honest estimate from a fantasy, so here are the line items, with real 2026 numbers attached to each one.

People are most of the bill

Start with the uncomfortable part. The median machine learning engineer in the United States earns $272,000 a year in total compensation, according to Levels.fyi. Salary guides for 2026 put senior base pay in a band from roughly $140,000 to $230,000, with total packages regularly clearing $300,000 once equity and bonuses land. And that’s the ordinary market. Frontier AI labs pay the same job title double or more, which keeps pulling the whole market upward.

For a budget, that translates simply. A small build team, say two senior engineers plus one product-minded generalist, costs somewhere between $60,000 and $90,000 a month once payroll taxes and benefits are included. Six months of that is $360,000 to $540,000 before a single customer touches anything. Agencies compress the number by putting fewer senior hours on your project than the pitch implied, which is a discount you usually discover later. Think of it like a home renovation. The materials are cheap. You are paying for the plumber.

The model itself is surprisingly cheap

AI vendors charge by the token, which is just a small chunk of text, about four characters. You pay one rate for the words you send in and a higher rate for the words the model sends back, the way postage depends on the weight of the parcel.

The current prices are public. Anthropic’s pricing page lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, mid-tier Sonnet models at $3 and $15, and the flagship Opus 4.8 at $5 and $25. OpenAI’s lineup spans a similar range as of July 2026, from about $1 in and $6 out at the cheap end of the GPT-5.6 family to $5 in and $30 out at the top.

Run the math on a real workload. A support product handling 10,000 conversations a month, at around 4,000 tokens per conversation, moves about 40 million tokens. On a small model that’s roughly a hundred dollars a month. On a flagship it’s low thousands. Prompt caching, which lets the vendor reuse the parts of your prompt that repeat, cuts those repeated inputs by 90 percent, and batch processing halves the price of anything that can wait a few hours.

But there’s a trap. Those numbers describe a chatbot. An agent, meaning an AI that performs multi-step tasks rather than just answering, can burn 20 to 100 times the tokens of a single chat reply, because it reads documents, plans, calls tools, checks its own work, and retries. Agent products are where API bills stop being a rounding error.

Infrastructure sits in the middle

Someone has to host the thing. You’ll pay for servers, a vector database (a filing cabinet that finds documents by meaning instead of by exact words), logging, monitoring, and a separate environment to test changes safely. For an MVP this is $500 to $5,000 a month on ordinary cloud services. For a production system it climbs to $10,000 to $50,000 or beyond, once you add redundancy, isolated environments per customer, and the log retention that auditors expect.

Data work is where budgets die

Here is the line item nearly every first budget misses. A restaurant meal takes twenty minutes at the table and hours in the prep kitchen, and AI products have the same shape. Before a model can answer questions about your business, someone has to find where the data actually lives, clean it, decide who is allowed to see which parts, and build the pipelines that keep it fresh. On real projects this quietly absorbs 25 to 50 percent of the engineering hours.

It never appears in the demo, which is exactly why it never appears in the quote. And it gets harder as companies get stricter about how AI is allowed to touch internal data, because every permission rule the security team adds is real engineering on your side of the bill.

Evals and safety are a real line item

An eval is an exam you write for the model, a fixed set of test questions with known good answers, run automatically every time anything changes. Without one, you find out about regressions from customers. The cost is senior engineering time to build the exam plus the inference cost of running thousands of test cases per release, and for anything customer-facing you should add deliberate attack testing, meaning people paid to make the system misbehave before strangers do it for free. Mature teams also budget for runtime defenses that watch the system while it operates, not just tests before launch. For an MVP, a few hundred test cases might do. For a regulated product, the eval suite becomes a permanent, documented asset that regulators can ask to see.

The run cost people forget

The build quote is not the price. It’s the down payment. After launch you pay for growing API usage, for monitoring and on-call, for prompts that need re-tuning as your data shifts, and for forced migrations, because vendors retire models on a schedule and Anthropic’s own pricing page carries a column of models already marked retired or deprecated. A sensible planning range is 30 to 50 percent of the original build cost, every year, indefinitely. An AI product is a boat, not a shed.

MVP versus regulated production

Now the totals. A focused MVP, one model API, one data source, no compliance surface, built by a small senior team in two to four months, lands between $50,000 and $250,000. Before committing even that, it’s worth an honest pass through the build versus buy question, because a $50-per-seat tool that covers 80 percent of the need beats a custom build on cost every time.

A production system in a regulated industry such as healthcare or finance is a different animal: $500,000 to $2.5 million or more in the first year, or 3 to 10 times the MVP. The multiplier isn’t padding. Every capability now needs an audit trail. Every release needs documented eval results. Lawyers review how data moves. A wrong answer stops being embarrassing and becomes a reportable event, so retrieval quality, meaning how reliably the system finds the right document before it answers, turns into an engineering program of its own. Narrow scope is the one reliable way to keep the multiplier near 3 instead of 10, which is the lesson Stripe’s compliance agents teach.

Where the money actually hides

Three places, and they’re predictable. Integration first: the AI is new, but it has to talk to a claims system installed in 2009, and connecting the two costs more than the model work. Compliance second: a SOC 2 audit, a HIPAA review, or a customer security questionnaire can each add weeks of engineering that produces nothing visible. Retrieval quality last, and largest. The demo answered beautifully from ten hand-picked documents. Production has to answer from 400,000 messy ones, and closing that gap is measured in months, not days.

So use the ranges as a lie detector. Any quote without a data-work line, an eval line, and a run-cost line is a demo budget wearing a product budget’s clothes. The vendors’ prices are public, the salary data is public, and the arithmetic above takes ten minutes. Do it before you sign, not after.