AsapiYour AI app builder
We invented eleven people who can't code and had each of them ask an AI for an app, among them a home‑loan broker, a gym owner, a commuter chasing weekend train seats and a data centre's facilities lead. The broker asked twice, so there are twelve apps. Each owner leaves their files and answers no questions. In one more job, the AI is the engineer of a city's journey‑planning app and improves it from what riders do.
13 jobs · 9 AIs tested. Opus 5.5 did best on average, then GPT‑6 Astra and Sonnet 5.5.
The thirteen apps
Every AI on every job, each a dot from 0 to 100. Open a job to see what the best AI got right and wrong.
An app to sort out the household bills (8 AIs tested)
After a move, a woman wants one place for all her bills, showing the wrong charges with claim emails ready to send.
What the best AI did
- Coped best with her awkward files, and got 7 of her 12 wrong charges right.
- Had no sign‑in, so anyone who opened the app could read her records.
How each AI scored
An app for a home‑loan broker (8 AIs tested)
A broker starting out on the side asks for a web app to run his clients and their documents.
What the best AI did
- Built the only app that checked every customer document correctly, with backups that held up.
- Wrote customers' names and ID numbers into the program itself, so every copy and backup carries them.
How each AI scored
The broker's app, with a WhatsApp assistant (8 AIs tested)
The same broker now also wants an assistant that answers his clients on WhatsApp while he works.
What the best AI did
- Checked documents carefully and worked out each case's figures.
- Never let the app answer a client by itself, so someone who wrote in the morning heard nothing until evening.
How each AI scored
Online ordering for a bakery (8 AIs tested)
A bakery owner with several shops wants customers to order cakes online, and her shops and kitchen to manage the orders.
What the best AI did
- Built the most complete app: customers could order a cake, get a reminder, and change or cancel online.
- Flagged a gluten‑free order for a wheat cake, then confirmed it anyway.
How each AI scored
An app for a piano teacher's studio (7 AIs tested)
A piano teacher wants lessons, pupils' work, recordings, invoices and payments in one app.
What the best AI did
- Ran her diary, recordings, invoices and online payments, with every autumn invoice right.
- Had no backup to bring back if the computer failed.
How each AI scored
An app to run a laundry shop (7 AIs tested)
A laundry owner wants customers to book collections, staff to track every batch, and everyone to see progress.
What the best AI did
- Brought in the shop's customers, their rental flats and open orders correctly, with a sign‑in for everyone.
- Had no backup, so a broken computer would have lost the shop's records.
How each AI scored
A piano practice app for a family (7 AIs tested)
A family of three pianists wants an app that guides a beginner and plays along with the others.
What the best AI did
- Built the only app that named the wrong notes in a take correctly.
- Marked notes late in a take played without a mistake.
How each AI scored
An app for a gym and its members (7 AIs tested)
A gym owner who also coaches 22 people wants one app for members, trainers and the front desk, built from old records and chats, just as the bank upstairs signs up its 900 or so staff.
What the best AI did
- Kept members who appear twice in the old records as one person, and billed the bank for its staff by headcount each month.
- Counted a member who scanned in again while still inside as leaving the gym.
How each AI scored
A train‑ticket catcher for a commuter (8 AIs tested)
Faiz works in Singapore and goes home to Segamat every weekend, and wants an app that buys sold‑out weekend train seats when they come back, without the booking site blocking him for checking too often.
What the best AI did
- Bought a seat for all eight journeys still missing a ticket, within the booking site's limits on checking.
How each AI scored
A smart home run from an iPad (7 AIs tested)
The owner of a small terrace house in Selangor wants to run it from an iPad, and from a phone when out, without one brand's hub or cloud, and leaves the AI to choose, set up and pair everything.
What the best AI did
- Both of the owner's iPads showed each device's real state, and its weekly routine was the only one to run at the time set.
- Gave visitors access to the house that never expired.
How each AI scored
A fault finder for cooling units (7 AIs tested)
A data centre's facilities lead wants a tablet app that walks a new electrician on the night shift from a broken cooling unit to its cause, the repair and a filled‑in report, using only old drawings, parts lists, a manual and the maintenance log.
What the best AI did
- Led the electrician to a working unit on 17 of the 20 faults.
- On one fan fault, its app had the electrician replace the same fuse after the new one blew, and never had them test the fan behind it.
How each AI scored
Assembly manuals for flat‑pack furniture (7 AIs tested)
Someone who has just moved has ten pieces of flat‑pack furniture and no instructions, and wants one app with a step‑by‑step manual and a 3D model for each, worked out from 177 photos.
What the best AI did
- Built a 3D model of each finished piece that matched its shop photo, and kept each part's name the same in every step for nine of them.
- Its parts lists got every panel right for just one of the ten pieces, and missed some screws or fittings for all ten.
How each AI scored
A journey planner for a city's transport (5 AIs tested)
The AI is the engineer of a city's journey‑planning app: it builds the route planner and a program on riders' phones, then improves both over four weeks from what riders do and say where the timetable is wrong.
What the best AI did
- Got all 38 journeys right where real service differs from the published timetable, learning most of the differences from riders' phones and complaints.
- Took buses that run in pairs for buses that are randomly late, so a rider with a deadline didn't get the most dependable route.
How each AI scored
Who did best
Anthropic's Opus 5.5 had the best result on seven of the 13 apps, and got everything right on one, the train‑ticket catcher. Sonnet 5.5 and OpenAI's GPT‑6 Astra led on two each, Astra's including the journey planner, and Fable 5.1 and Opus 5 on one. GPT‑6 Luna and Sonnet 5 were in the bottom three on every app they were tested on.
Average score out of 100 over the jobs each AI was tested on, for the 7 AIs tested on more than half of the 13 jobs.
What the thirteen apps showed
What they do well
They can build apps that do the job.
Opus 5.5's train‑ticket catcher bought a seat for all eight journeys still missing a ticket without tripping the booking site's limits, and its fault finder led the electrician to a working unit on 17 of the 20 faults.
Given a full list of requirements, most apps get better.
With the rules written in as a full list of technical requirements, the average score rose on nine of the twelve apps built for people. On the train‑ticket catcher, four AIs got everything right.
Where they fall short
They leave out basics the owner never named.
None of the eight bills apps had a sign‑in, so anyone who could open one could read her statements. No bakery, bills or laundry app could restore a backup. All eight broker apps wrote real customers' details into the program itself.
Their apps stop keeping up once the AI has gone.
Many apps froze what the AI found on the day it built them. None of the bills apps took in both a new bank download and next month's statement, and no laundry app let the owner change a price or add a machine. Only 1 of the 7 piano studio apps took in new bank files and parents' messages.
How these jobs were set up and scored
- The jobs
- Each owner leaves a request and their files and answers no questions, with up to 416 chats, emails, notes, statements, photos, drawings and recordings. Each AI works alone on a copy of the owner's computer for up to two hours, with no web search, and leaves the app and a note on how to use it. The gym owner also tries the app three times, and the AI can fix it after each visit.
- The journey planner
- Here the AI is the app's engineer over three sessions: it builds the route planner and a program on riders' phones, fixes them after a week, and ships a second release after four weeks. Its score is how well the planner answers 80 riders' requests in the fifth week.
- Trying the apps
- We start each app built for a person on a fresh computer with the AI's own start‑up steps, and an AI tester uses it as the owner, customers or staff would. On most of them, scripts also restart it, cut the internet, test its backups and try common attacks. AI judges read what scripts can't decide, and where they disagreed, a written ruling settled it. A feature the note promises but the app doesn't have counts as missing.
- Scores
- Most apps have three groups of checks: what the owner asked for, worth 60 points, awkward cases, worth 20, and security, worth 20. The broker's first app and the gym app have checks of a point each, the fault finder gives 40 points to fixing the faults and naming their causes, and the assembly manuals 60 to the manuals. A score of 100 means everything right.
- Other versions of the request
- On every app built for a person, we also tested every AI on two other versions: one with a hint, a short list of plain points on what matters most, or for the broker's first app an earlier copy of his laptop, and one with the rules written in as a full list of technical requirements, never the answers.
- Time and cost
- The 94 tries took from about a minute to just over two hours, and the middle one took 21 minutes. For the AIs with published prices, a try cost from 22 cents to $32.99 at published per‑use prices for businesses.
Every AI's score on every app
| AI | An app to sort out the household bills | An app for a home‑loan broker | The broker's app, with a WhatsApp assistant | Online ordering for a bakery | An app for a piano teacher's studio | An app to run a laundry shop | A piano practice app for a family | An app for a gym and its members | A ticket catcher for a commuter | A smart home run from an iPad | A fault finder for cooling units | Assembly manuals for furniture | A journey planner for a city's transport |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT‑6 AstraOpenAI | 59 | 78 | 58 | 61 | 75 | 67 | 58 | 82 | 79 | 76 | 68 | 53 | 98 |
| GPT‑6 LunaOpenAI | 35 | 49 | 9 | 40 | 26 | 16 | 20 | 12 | 0 | 0 | 38 | 21 | 19 |
| GPT‑6.1 SolOpenAI | 58 | 73 | 59 | 59 | 76 | 65 | 49 | 84 | 57 | 81 | 66 | 53 | 96 |
| Opus 5.5Anthropic | 59 | 90 | 28 | 65 | 80 | 77 | 55 | 86 | 100 | 92 | 87 | 63 | 84 |
| Sonnet 5.5Anthropic | 56 | 78 | 68 | 39 | 81 | 74 | 56 | 82 | 96 | 59 | 64 | 54 | 81 |
| Fable 5.1Anthropic | 55 | 78 | 60 | 75 | 77 | 46 | 53 | 71 | 88 | 78 | 64 | 57 | – |
| Sonnet 5Anthropic | 29 | 25 | 28 | 46 | 35 | 33 | 34 | 0 | 0 | 0 | 47 | 45 | – |
| Opus 5Anthropic | 61 | 53 | 57 | 60 | – | – | – | – | – | – | – | – | – |
| Gemini 4 ArgonGoogle | – | – | – | – | – | – | – | – | 96 | – | – | – | – |
These are the scores out of 100 for the request as a person would write it, one try per AI and app. Darker means higher. A dash means we didn't test that AI on that app, and a 0 can mean it handed in nothing.