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 job site is not the keynote.
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 screen keeps moving, so people think the work is 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 is a stronger tool, not a free upgrade in quality, and not a formula for calendars.
More drafts is not more software shipped
Typing speeds up. Commits pile. Test files multiply. 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 people — matching the spec, reading permissions, fixing APIs glued the wrong way, deciding who presses go-live.
Industry notes that almost everyone uses the tools while pull-request throughput only inches up land in the same place as our quotes: tool fees, tokens, a senior person watching the diff. If the “saved” afternoon is spent undoing a permission the model invented, the invoice does not shrink.
A jam we see often: the payment contract is not back, and the screen already says the order went through. The model is happy to draw the button and even invent a success line. The week you can actually launch is still waiting on the bank, 3-D secure, and where money goes when it fails. Those days do not vanish because drafts appear faster. Role names the model can guess; which report each role may open, only your staff know — and that table often takes three passes.
It can run at 3 a.m. The name on the release is still a person
A model can keep drafting overnight. The build that ships still needs someone who owns the scope, someone who runs the tests, someone who checks personal data did not land in the wrong place. Swapping an engineer for “a person watching the screen” is not free. If nobody understands the system, the mistake goes to production and the customer finds it first.
No serious trial supports crushing a full systems job to one-twelfth of a year. A lot of that year is not typing: moving old data, waiting on contracts, changing three fields after acceptance. Keynotes do not film those queues.
We still quote on scope and risk. AI speeds repeated stretches and first drafts. We do not sell an “AI timeline” discount. Ballpark numbers stay on the services page and in a call, not on a model launch stage.
FAQs
Do you still use AI to write code?
Yes — drafts, repeated refactors, looking up APIs. A person still owns logic, permissions, tests, and the spec you signed. Tools do not auto-halve the quote.
What if a small change really ships in a day?
Copy and extra fields can be fast. Payments, legacy systems, access control, and acceptance are not the same job.
Why not generate everything first and only review at the end?
Review is not the last five minutes. Permissions and payments written into the wrong place become production incidents. Someone has to be in the spec, not only in QA.
Bring a spec, talk timeline Services
Sources: METR 2025 RCT; METR Feb 2026; CEPR on writing vs shipping.