How much does custom software cost now? It depends on the size of the project. One firm puts a small first version at $15,000 to $50,000. Another says a large system can cost $400,000 or more. AI agents now write much of the code, so the coding costs less. Planning, review and security don't shrink the same way, so the total falls by less than the coding does.

That breaks the old way of pricing a quote, which was hours times a rate. It doesn't change where most buyers lose money. They still pick the cheapest seller.

I sold and delivered custom software for nine years at Upstack Studio. Here's how I'd read a software quote today.

What does custom software cost in 2026?

Here's what two agency guides that rank for this question say. Each figure is that firm's own estimate, not a market average.

Project type Hipro Technologies (August 2026), quoted in Selleo's guide SOLTECH (2026 guide)
First version or small app $15,000 to $50,000 for a web or mobile first version (MVP) $100,000 to $200,000 for a small app
Customer app or SaaS platform $50,000 to $180,000 for a SaaS platform $200,000 to $400,000 for a medium-sized app
Large or complex system From $150,000 for an enterprise system that connects to other systems $400,000 or more for a large or enterprise app

Sources: Selleo's guide, updated October 2026, and SOLTECH's 2026 guide.

Even a first version has no single price. Selleo also quotes a second firm, Horizon Dynamics (February 2026), at $40,000 to $120,000 for one. Each firm defines the project its own way. Use these as rough starting points.

How does AI change those numbers?

The sources don't agree on how much AI saves:

  • SOLTECH says that in some cases, AI can cut the development work by roughly 20–30%.
  • Selleo says AI tools don't give a reliable saving across a whole project.
  • A Harvard study found engineers spend 30–40% of their time writing code. The rest goes on planning, review and testing.

Put together, that means AI makes the coding cheaper. Coding is only part of the job, so the whole project falls by less. Planning and review still need people.

I think the cost of custom software will change for good. Build time is much shorter now, and the old time estimates no longer hold.

What should custom software cost now?

There's no one right number. The ranges above are a fair place to start. What should change is what the money pays for.

My view is that you should pay for judgment, review and results, not hours of typing. A fair quote now does two things:

  • It shows where AI cuts the hours.
  • It prices the result you get, not the time it takes.

If a quote still counts every hour of coding at the old rate, you're paying for work the AI now does. If it's far below the ranges above, ask who does the planning, review and security work.

What drives the price?

These are the main things that move a quote up or down. Both guides name most of them.

  • Scope: how many features you need, and how complex they are.
  • Connections to other systems: each one your software has to talk to adds work.
  • Design and platforms: a web app costs less than a web app plus iPhone and Android apps.
  • Old data and old systems: moving data over takes time.
  • Security and testing: the checks that make it safe to use.
  • Who builds it: SOLTECH lists $50 to $250 or more an hour for US developers, and $20 to $80 in Eastern Europe and Asia.
  • Upkeep after launch: hosting, support and fixes.

AI changes the cost of the build. It doesn't remove any of these.

What got cheaper: the build

Before AI, a project needed people for each layer:

  • the back end
  • the front end
  • the screens

Each layer took weeks of hands-on coding.

Now AI agents can do much of that work on every layer. So the people on the project spend their time in a new way. They direct the AI and review what it made. They get it to fix things before anything ships.

Real company data shows more code, but not much more finished work. Two Harvard researchers, Fiona Chen and James Stratton, studied 720,542 engineers at 655 firms (paper version of 6 November 2026). A year after firms started using AI coding agents:

  • engineers wrote 44% more lines of code
  • tasks closed rose 13.7%
  • larger projects completed rose 18%

The researchers say none of these rises is firmly statistically significant. AI sped up the writing. The other steps became the bottleneck.

Where the money goes now

↓ Going downWriting the codeBack end, front end and screens, now done largely by AI agents
↑ Going upJudgmentDeciding what to build and what to leave out, directing the AI, reviewing its work, security and system design

Typing got cheaper. Thinking and checking cost more.

My view of where custom software cost is moving. The Harvard study found engineers spend 30–40% of their time writing code.

What got more expensive: judgment

What you pay for now is the brain. Someone has to decide what to build and what not to build. Someone has to set the product plan and point the AI agents in the right direction. That person decides whether you get a result or just a lot of code.

Planning matters more than it did before AI. An AI agent needs clear direction before it can start. Give it vague direction and you get a fast build of the wrong thing. As I wrote in my build vs buy guide, the expensive mistake today is a successful build of the wrong thing.

Checking the work is a judgment cost too. In the Harvard study, after firms adopted AI agents:

  • code took 56% longer to go from review request to approval
  • each change drew 39% more review comments

More code arrived, and skilled people had to read it.

Why developer skill still matters

Let AI agents run loose and you can get quite far. Then the build hits the hard parts:

  • security
  • architecture, or how the system fits together
  • growth, or whether it holds up as more people use it

Those need people who know what they're doing.

Security shows the gap. Veracode tests AI models on coding tasks where there's a safe and an unsafe way to write the code. In its October 2026 test of 18 models, the average model wrote safe code on 62% of the tasks. The best models reached 73%. So even the top models wrote unsafe code on more than a quarter of the tasks.

A developer who knows security and how to design a system that grows now spends fewer hours typing. Each of those hours is worth more. It's the hour that catches what the AI got wrong.

Should you still pay for time and materials?

Time and materials means you pay for the hours worked. That made sense when hours were most of the cost. When AI does most of the build, the hours stop telling you much about the value. Pricing should follow results.

Here's how to put that into practice:

  1. Write down what "done" means: what the software must do, and how you'll check it works.
  2. Ask for a price for that result, or a price for each clear stage.
  3. Agree up front how changes get priced. A fixed price only covers the agreed work.

A results-based quote needs a clear brief. That's one more reason the planning is worth paying for.

Where buyers still lose money

Where buyers lose money hasn't changed much. They trust the seller with the lowest price and try to save money on the build. That was a bad trade before AI. It's a worse one now.

When AI does the build, the quality of what you get depends on who directs it. A cheap team that lets the AI run loose can show you a working demo fast. Security, scale and upkeep are where it costs you later.

So who you pick matters more than ever. That goes for a developer, a product manager or an agency.

How to choose who builds it

I'd judge a software partner on four things. I'd skip most questions about frameworks and coding languages.

Four checks

  1. 1How they planand what they chose not to build
  2. 2How they treat clients
  3. 3How they use AIand who checks its work
  4. 4How they stop AI slopgetting into your product

Skip the framework quiz. Judge the people who direct the AI.

Four checks for choosing who builds your custom software.
  1. Their portfolio and how they plan. Ask them to walk you through a past project. What did they decide to build, and what did they leave out? Why?
  2. How they work with clients. Watch how they run your first meetings. Ask how they handle a change of plan or bad news.
  3. How they use AI, and their checks. Which parts does the AI write? Who reviews its work, and what testing runs before release?
  4. How they keep AI slop out of the product. AI slop is work that looks finished but is generic, unchecked or wrong. Ask them to show you how they catch it.

Want more to ask before you sign? Use my list of questions to ask an AI or software vendor.

Questions owners ask next

Does AI make custom software cheaper?

The coding part, often yes. The whole project, by less. Ask each vendor where AI cuts their hours, and check that it shows in the price.

Can I build it myself with AI now?

You can get far on your own. Before you start, decide who owns security and upkeep once it's live. My build vs buy guide walks through that call.

What does it cost after launch?

Custom software keeps costing money after launch. The cost of building fell. The cost of owning what you built did not. Over the first year, count:

  • fixes and updates
  • hosting
  • your team's time

Selleo notes there's no fixed yearly percentage that fits every product.

How do I set a budget before asking for quotes?

Start with what the problem costs you each month. Lost sales and staff hours both count. This guide to where AI goes first shows how to put a number on it.

Before you ask for a quote

  1. Write down the problem, what it costs you and what "done" looks like.
  2. Send the same page to each seller.
  3. Choose the seller with the best judgment, even if they aren't the cheapest.