AI Software Development Cost: What You Pay to Build It, and What You Pay Every Month After

AI Software Development Cost: What You Pay to Build It, and What You Pay Every Month After

Ask what AI software development cost looks like in 2026 and you get a range. The published guides land somewhere between $15,000 and $2 million, and every one of them is telling the truth. That is the problem. A weekend proof of concept and a regulated enterprise platform sit inside the same sentence, so the range prices nothing you can plan against.

There is a more useful way to hold the number. It starts by admitting there are two of them: what you pay a team to build the thing, and what you pay every month afterwards to keep it answering. Almost every proposal covers the first and goes quiet about the second.

The build number is hours times a rate

That is the whole formula. A vendor who gives you a figure before scoping has picked something that sounds survivable, and the scope gets bent later to fit it.

An honest estimate names hours per line item. The line items are boring:

• Deciding what the model actually decides. Not "add AI to support" but "classify the ticket, draft a reply, hand anything about refunds to a human." A vague decision is the most expensive sentence in the document.
• Data work: getting your content into a form the system can retrieve from. Extraction, cleaning, chunking, permissions. On most projects this is the largest single block, and the one clients least expect.
• The integration surface. Every system the feature reads from or writes to. Two integrations is a project. Six is a different project.
• An evaluation harness — a set of real inputs with known-good answers, plus a way to run it on every change. Teams skip this to save two weeks, then spend two months arguing about whether the thing got worse.
• The product around the model: auth, admin screens, rate limits, logging, an audit trail, a way for a human to override. Ordinary software. Usually more hours than the model work.
• The accuracy bar. Getting to roughly-right is fast. Every step after that costs more than the step before it.

Two of those are where budgets die. A demo that is right 80% of the time takes weeks; something you would put in front of a paying customer unsupervised is a different order of work, and no honest quote hides that behind a single figure.

The monthly bill is arithmetic you can do before you sign

You can nail this part down precisely. Hardly anyone does it in the sales conversation.

Take a support assistant handling 10,000 conversations a month, six model calls per conversation, so 60,000 calls. Each call sends around 4,000 input tokens — your instructions plus whatever documents got retrieved — and returns about 400. That comes to 240 million input tokens and 24 million output tokens a month.

Against Anthropic's published list prices, read on 30 August 2026:

• Haiku 4.5, at $1 per million input tokens and $5 per million output: $360 a month.
• Sonnet 5, at $2 and $10: $720 a month.
• Opus 5, at $5 and $25: $1,800 a month.

Same workload. Five times apart. The difference is one line in a config file, which makes model choice a pricing decision rather than an engineering preference, and it deserves revisiting per feature instead of being picked once for the whole product.

Now the second thing. Most of those input tokens repeat: your instructions and your retrieved documents are identical call after call. Prompt caching bills that repeated context at a fraction of the rate, $0.20 per million tokens on Sonnet instead of $2. If three quarters of your input is stable, the Sonnet row above drops from $720 to roughly $396. There is a one-off write cost per cache window. It is real and it is small next to the saving.

Batch processing takes another 50% off anything that does not need an answer this second.

Then scale it. Ten times the traffic on Sonnet, uncached, is $7,200 a month. Call it $86,000 a year. If your build came in at $60,000, the running cost overtakes it inside the first year — and that is the sentence missing from most cost guides, the one that should be shaping your architecture rather than surprising you in month nine.

Tokens are not the whole bill. Add hosting, a vector database, logging and evaluation runs, monitoring, and the human hours someone spends reviewing what the system got wrong. Budget for those or meet them as a surprise.

One more multiplier. Agent-style features that plan and use tools do not make one call per task; they make one per step. A twelve-step agent is twelve calls, and your cost per task moves accordingly.

What actually moves the build number

Four things, in the order they bite.

How ready your data is. Clean, permissioned content shortens a project more than any framework choice anyone will argue with you about.

How many systems it touches — every integration carries auth, rate limits, edge cases, and other people's outages.

How wrong it is allowed to be. The distance between 80% and 97% is where most of the money goes, and it is not a straight line: the last few points can cost more than everything before them, because they are bought with evaluation data, retrieval work, and fallbacks rather than with a better model. Decide that bar in writing before anyone estimates against it.

And what happens when it is wrong. A bad product recommendation is a bad day. A wrong dosage or a wrong price is a different budget, with review workflows and audit trails attached.

How to buy it without getting burned

Pay for a short scoping phase and treat its output as the deliverable: the decision the system makes, the data inventory, the integration list, the accuracy bar, an evaluation set. A vendor unwilling to sell you that on its own is a vendor planning to discover your scope on your budget.

Then split the contract. Genuinely known work — screens and integrations — can be fixed price. Research runs on time and materials with a cap and a weekly burn report, and model behaviour is always research. We have written before about choosing between time and materials and fixed price and how to run a time and materials engagement without losing control of it.

Ask for the monthly running cost in the proposal, with the token arithmetic shown. If a vendor cannot produce that, they have not thought about your product past launch day.

We run our own AI product, so the bill lands on us every month and we watch that line the way our clients watch theirs. It changes how you scope. You stop reaching for the biggest model by default and start asking what each call is actually worth.

Pricing a build right now? Our AI development services page covers what we take on, and our post on AI software development services covers how we run the work. Send us the scope. You get hours per phase and a monthly running estimate back, and you are free to take both to someone else.