How Much Does Custom Software Cost? Estimator built on 125 real projects

What 125 real products cost

Reality check always beats estimates.
me, on LinkedIn, about the chart below

Atomic Object, a US product studio, publishes the cost of products it built over five years. It is the best public dataset I know: finished projects, not quotes. Their typical rate on Clutch is $150 to $199 an hour, so the dollars below are premium-agency dollars; the hours behind them are what this estimator reuses.

Discovery and designMobile appWeb appEmbeddedDesktop Data: Atomic Object, “The Cost of Quality Software”, products built over the last 5 years, retrieved Oct 17, 2026. Chart redrawn from their data.

Cost percentiles by type (USD)
Type	n	p25	median	p75	p90
Web app	58	$152k	$311k	$882k	$1.4M
Mobile app	23	$96k	$210k	$412k	$1.2M
Discovery and design only	36	$28k	$46k	$100k	$162k
Desktop or embedded	8	$144k	$229k	$312k	$527k
All	125	$73k	$162k	$409k	$1M

Is Atomic Object an outlier? Three more datasets

Atomic’s numbers come from one studio. Research datasets in the PROMISE repository record effort in person-hours for finished projects elsewhere: a Finnish commercial bank (62 projects, 1985–1993), Computer Sciences Corporation (145 projects) and the China dataset (499 projects). They are old, so they say nothing about today’s prices. They say a lot about the shape: in every set the 90th percentile is many times the 10th, and a few projects dwarf the rest.

Effort in person-hours by dataset
Dataset	Projects	p10	p25	median	p75	p90	p90 ÷ p10
Atomic Object (cost ÷ $175)	125	216	420	924	2,339	5,906	27.4×
Finnish bank (Maxwell)	62	1,113	2,389	5,190	9,308	14,947	13.4×
CSC (Kitchenham)	145	492	874	1,557	2,895	5,417	11×
China	499	351	704	1,829	3,827	8,708	24.8×

The CSC set also records the first estimate of each project. Half of the projects finished between 0.79× and 1.07× of it, 14 percent overran it by more than 25 percent, and the worst took 5.9× the estimate. That is a mature outsourcer estimating work it does every day, so read it as the best case, not the typical one.

Bigger benchmarks exist but cannot be republished for free: QSM analysed 2,231 projects completed since 2000, and ISBSG sells a repository of thousands. Price calculators and “average app cost” articles are not data: they rarely say how many projects stand behind a number.

Why the range is so wide

Not because estimators are bad. At the idea stage nobody, the client included, knows yet what will be built. Steve McConnell’s cone of uncertainty puts the spread at 0.25× to 4× of the estimate at the initial concept, 0.5× to 2× once the product definition is agreed, 0.67× to 1.5× with requirements written, 0.8× to 1.25× with the UI designed. And that is the best case.

Surveys of real projects agree on the direction: 60 to 80 percent of software projects overrun effort or schedule, by 30 to 40 percent on average (Moløkken and Jørgensen). In 1,471 IT projects the mean cost overrun was 27 percent, but one in six overran by about 200 percent (Flyvbjerg and Budzier). The average hides the tail, and the tail is what sinks a startup. At the large end, McKinsey and the University of Oxford found, across more than 5,400 IT projects, that projects with budgets over $15 million ran 45 percent over budget and 7 percent over time and delivered 56 percent less value than predicted. Every extra year of duration added 15 percent to the overrun. Short projects are not just cheaper; they are more predictable.

Fixed price, time and materials, or a fixed budget?

The contract decides who carries the uncertainty above. None removes it.

Contract aspect	Fixed price	Time and materials	Fixed budget, flexible scope	How I work
What is fixed	Scope and price	Rate and quality	Budget and quality	Price per slice, or a weekly cap
What gives	Quality and goodwill, or change requests	Cost and timeline	Scope, re-prioritized weekly	Scope of the next slice
Who carries the risk	Vendor, so the price includes a buffer	Client	Shared	Shared, in small steps
Fits when	Scope is small and truly known	Exploration, ongoing work, trust	A product with a budget and a goal	Founders who want to see working software every week

The fixed budget, flexible scope model is Atomic Object’s, which they describe as “fundamentally a T&M contract with a Not To Exceed limit”. On fixed price: in 93 offshore projects, vendors earned more on time and materials on average, but preferred fixed price for larger, longer projects, where knowing more than the client lets them price in a margin (Gopal and Sivaramakrishnan, 2008). Nobody publishes a reliable size for that buffer, so this page does not invent one.

AI coding agents: my multiplier, and why it is a slider

Agents are how I work today. They write most of the code, run the tests and propose fixes; I decide what to build, design the parts that matter and review every change. The market rows in the table above come from projects built before agents, so they get no discount. My row gets the multiplier you set with the slider, 5× by default.

The public evidence is moving fast and is still thin for one person. When Microsoft rolled out Claude Code and Copilot CLI to tens of thousands of engineers in early 2026, adopters merged about 24 percent more pull requests than they would have otherwise. Anthropic reports a 67 percent increase in merged pull requests per engineer per day after adopting Claude Code, which is a vendor measuring itself. Google’s CEO said 75 percent of new code at Google is AI-generated. In the 2026 Stack Overflow survey, 66 percent of developers who use AI use coding agents. Earlier controlled trials timed single tasks: 21 percent less time at Google, 55.8 percent faster on a small greenfield task, and 19 percent slower for experienced open-source developers in early 2025. METR now calls its own newer study a bad proxy and very weak evidence, most likely a lower bound. Nobody has a controlled measurement of 5× to 10× for one senior with agents. That number is my experience, which is why it is a slider and not a fact.

What the multiplier depends on

Agents speed up the part of the work that is typing, testing and searching. Everything else runs at human speed, so the overall gain is capped by the slowest part (Amdahl’s law, applied to a project). In a regulated company, such as a medical device maker working to IEC 62304, every change also needs validation, review records and a sign-off; a first draft written ten times faster can end up maybe 1.5× faster overall. A low appetite for risk lowers it further, and that is often the right call.

What lowers and raises the AI multiplier
Factor	Pushes it down (1–3×)	Pushes it up (7–10×)
Regulation	Medical, finance, children’s data: validation, audits and sign-off for each change	No compliance gates beyond good practice
Decisions	A committee; approvals take days	One owner who answers within hours
Codebase	Legacy without tests, behaviour nobody remembers	A new product, or a codebase with tests and CI
Environments	No production-like data or devices, manual deploys	Preview environments, automated deploys
Scope	A moving target, specs written after the fact	A clear slice with acceptance criteria
Risk appetite	Every change justified on paper	Ship small, measure, roll back fast

With great power comes great responsibility

With AI, people can also smash their stupid heads ten times faster, as I keep saying on LinkedIn. Veracode’s 2026 report found that only 56 percent of AI-generated code passes security tests, virtually unchanged from a year earlier, even as models got better at everything else. In 2025 an agent deleted a production database during a code freeze and then fabricated data. DORA 2025 calls AI a mirror and a multiplier: it amplifies good engineering practice and exposes weak practice.

I review every change; agents run the tests before I see the code.
Hooks block risky commands, and secrets never reach the model.
Nothing goes to production without a human decision (human in the loop).
I would rather ship a smaller slice than slop. More in the glossary.

Method and limits

This is reference class forecasting: start from what similar finished projects cost, not from a feature list (Flyvbjerg; McKinsey recommends the same check).
Dollars from Atomic Object’s chart are converted to hours at $175, the middle of their Clutch rate. The hours are then priced at other rates.
Your answers to step 2 move the project from the 10th to the 90th percentile of its type. That mapping is my heuristic from years of estimating, not data.
Cross-cutting needs (languages, offline, accessibility) cost more than they look, because they touch every screen. That is why they weigh more than single features.
The dataset is one US studio’s portfolio: complex products with discovery, design and QA included. A landing page or a no-code prototype costs far less.
Rates: Clutch Poland $50–99/h (2026 guides), Accelerance Europe senior $64–76/h (2026), my rate 150–200 PLN/h. Currency at the NBP rate of 2026-10-16 (1 USD = 3.8827 PLN).
My row divides hours by the AI multiplier from the slider; the market rows do not, because their data predates agents. The multiplier is my judgment, not data. Set it lower if your project looks like the left column of the table above.

Sources

Questions

Why not count features and multiply by hours?

Because nobody publishes credible hours per feature, and the features you list today are not the ones you will ship. A distribution of finished projects already contains the surprises.

Is the number on this page a quote?

No. It is a range to plan a budget and a conversation. A quote needs a scope, and a good scope usually cuts the first release in half.

Why is your rate shown next to agencies?

Because it is my tool, and I would rather show the comparison than hide it. The duration column matters as much as the price: one person cannot parallelize like a team of four.

Is 5× realistic?

For a new product with fast decisions, it is what I see in my own work. For regulated or legacy work it is lower, and I will tell you so on the call. That is why 5× is the middle of the slider, not the top.

How much will your software cost?