Does AI coding really go faster and cheaper — and finish a year of work in a month?
The folders are already closed. Then, lightly, as if asking about the weather: “Can’t one person plus the tools do it now?” We have heard that sentence a lot. Keynotes show code appearing by itself. A year of work looks like it might fit in a month. The system we are actually building is not that video. The spec still has to match, the old data still has to move, and acceptance still goes back and forth. None of that disappears because code grew itself on a stage.
The stopwatch and the feeling disagree
METR ran a trial in early 2025: experienced developers, real issues on their own repos, AI allowed on a random half. The clock said they took about 19% longer. Afterward they still thought they had been 20% faster. The number got over-quoted. The ordinary part is this: the model keeps typing on the screen, so the person writing the code thinks the ticket is getting finished. The clock measured how long the ticket took to finish, not how long the screen was moving.
Tools changed. METR’s February 2026 note leans toward a speedup and immediately says fewer people will sit a test with AI turned off, so the sample is bent. They do not treat that figure as settled. Product accounts still push Codex desktops that drive the computer and split subtasks (see @thsottiaux). That means the coding tool got stronger. It does not mean the system you hand over is automatically more correct, and it does not mean you can divide a year of work by twelve and call that the new schedule.
More drafts is not more software shipped
The model and the engineer both type faster. The repository gains more commits and more test files. What increased is files, not features already released for a customer to use. CEPR’s write-up of a large GitHub study is blunt: each generation of tools lifts “code written” a lot; releases do not rise in step. What sits in the middle is work a person still does: match the program to the spec you agreed, see whether each role can read data it should not, fix an API wired to the wrong endpoint, and have someone willing to press publish on launch day.
Industry notes that almost everyone uses the tools, while the pull requests people review only inch up, land in the same place as our quotes. The tool’s monthly fee, the tokens spent calling the model, and a senior engineer reading each diff all sit on this job. They do not vanish from the quote because “there is AI.” If the afternoon you saved is spent entirely on fixing a permission the model wrote wrong, the invoice the owner receives does not get smaller.
A jam we see often: the payment contract is not back, and the screen already lets someone press “place order.” The model is happy to draw that button and even invent the line “order placed.” The week you can actually launch is still waiting on the bank’s merchant contract, on whether 3-D Secure is live, and on whether a failed charge returns to the cardholder, sits as an authorisation, or has to be matched by hand. Those waiting days do not disappear because the draft of the button appeared faster. The permission table is the same kind of wait, and it often takes three passes with you. The model can guess role names such as admin or staff. Which role may open which report, only people at your company know. We do not put the guessed table live.
It can run at 3 a.m. The name on the release is still a person
A model can still be drafting at three in the morning. The build that ships still needs a person who owns this scope, a person who runs the tests, and a person who checks that a customer’s personal data was not written into the wrong environment, for example the test database and the production database swapped. Replacing an engineer with someone who only watches the model and does not understand the system is not free. If nobody understands the system, a wrong permission or a wrong write goes into production, and the customer hits it before your engineering team does.
No serious controlled trial supports compressing a year-long systems job into one month. A lot of that year is not typing: how old data moves into the new system, when the contract comes back, and the owner sending a form back at acceptance to rename a field or make it required. Keynote videos do not film those waits.
We still quote on scope and risk. AI speeds repeated blocks of code and the first draft. We do not use “we used AI, so the schedule shrinks and the quote is discounted” as a line. Figures for a brand site or a systems job stay on the services page and in the scope you confirm in a conversation, not in a video from a model launch.
FAQs
Do you still use AI to write code?
Yes. We use it to draft, to rewrite repeated blocks of code, and to look up how a function is called. Before launch a person still reads the logic, the permission checks and the tests against the spec you confirmed. The quote does not halve itself because the job uses that tool.
What if a small change really ships in a day?
It depends what is changing. Copy on a page, or one extra form field, can be done in a day. Connecting payments, connecting an old system, agreeing permissions, and running acceptance are not that job. Do not estimate them at the speed of a copy change.
Why not generate everything first and only review at the end?
Review is not the last five minutes before launch. If a permission, personal data, or a payment call is written into the wrong environment, what you undo is live orders and access, not a draft file. A person has to be in the spec from the start, not only in the final QA pass.
Bring a spec, talk timeline Services
Sources: METR 2025 RCT; METR Feb 2026; CEPR on writing vs shipping.