Business operations · Task 3 of 7
A bakery owner's ordering app
Build a working ordering app for a bakery owner from her laptop, where everyday records reveal what customers, shop staff and the kitchen need.
The task
A bakery owner wants an online ordering app that works for her. She leaves her laptop, with business records, customer complaints, staff conversations and an existing web form. Nobody answers questions. The agent must build the app and leave a short note that lets her use it.
The prompt only asks for an app and a note; the real test is recognising the work behind an order. Customers change their minds, shops need separate views, the kitchen needs reliable lists, and the owner needs to recover her data and control access. Past mistakes and competing suggestions make a plausible first draft easy to get wrong.
What it tests
Working out the real needs
The request is short, while the operational needs emerge across ordinary business records.
Weighing conflicting suggestions
An old form and informal advice can disagree with how the business actually works.
Connecting whole workflows
An accepted order must stay correct through changes, collection, kitchen work and owner reports.
Shipping usable software
The app must start on a fresh machine and support people through its own pages.
Challenges
1Changes after booking
Different people need different powers when an order changes late.
2Work across shops
Staff and kitchen views need useful information without exposing unrelated customer records.
3Quiet background work
Reminders, backups and recovery need to work after the agent leaves.
4Ordinary awkward input
Real names, telephone numbers and customer messages expose assumptions that neat examples miss.
5Protection without obstruction
Security controls must stop abuse while letting normal customers complete their work.
Results
| Model | Standard | Hinted | Told everything | Time (standard) | Cost (standard) |
|---|---|---|---|---|---|
| Fable 5.1 | 75% | 77% | 99% | 16 min | $8.35 |
| Opus 5.5 | 65% | 82% | 52% | 11 min | $2.95 |
| GPT-6 Astra | 61% | 38% | 96% | 6 min | $2.11 |
| Opus 5 | 60% | 72% | 68% | 18 min | $7.31 |
| GPT-6.1 Sol | 59% | 61% | 94% | 5 min | $0.31 |
| Sonnet 5 | 46% | 63% | 90% | 8 min | $1.45 |
| GPT-6 Luna | 40% | 56% | 80% | 3 min | $2.78 |
| Sonnet 5.5 | 39% | 36% | 99% | 6 min | $1.06 |
One run per model and setup. 100% = everything right (or, for a plan scored in money, the best plan); 0% = the obvious answer or worse.
Setups: Standard, the job as asked; Hinted, plus one hint; Told everything, every rule or fact given up front: the control that shows the task is fair.
What models got right and wrong
- Fable 5.1 led the standard setup with 75.2%. With everything told, it missed only a late staff change that its customer form still blocked.
- Fable 5.1, Opus 5.5, GPT-6 Astra, Opus 5, GPT-6.1 Sol, Sonnet 5, GPT-6 Luna and Sonnet 5.5 all left gaps in owner tools or handover on the standard setup.
- Sonnet 5.5's standard and hinted apps refused their own browser submissions. With everything told, it completed the ordering and unusual-input work but let too much automated traffic create orders.
- Opus 5.5's hinted app outscored its told-everything app. The latter rejected an ordinary kind of mobile number, stopping many later workflows from starting.
- GPT-6 Astra's hinted app moved the working date away from the supplied date, although its own tests supplied a different clock.
- GPT-6.1 Sol handled the unusual-input and security work with everything told, but its app hid a paused product and missed parts of the owner's handover.
Test environment
- Kind
- Simulated bakery business: the owner's laptop is a folder of files; the agent builds a working web app
- Agent's interface
- A terminal in a fresh Debian Docker container, driven by Claude Code or Codex CLI; Python 3.13, Node 22, document readers, Playwright and Chromium are available
- Documents
- 87 files read-only at
/home/user/morwenna/: 17 PDF, 17 DOCX, 11 XLSX, 34 EML, three TXT, two HTML, one Python, one CSS and one JavaScript file; these include the existing web form - Applications
- A local SMTP relay at
127.0.0.1:2525records mail without sending it outside the container; LibreOffice, Pandoc and document-reading libraries are available; there is no mail client or chat client - Web
- None. Internet access is available for package installation; the agents' own web search and fetch tools are off
- People
- Nobody answers. The owner, staff and customers appear only in saved documents, emails and chat exports
- Time
- Monday 12 April 2027, 11:00 in Europe/London. During grading, the app's clock ticks from that time, freezes for exact notice checks, and moves forward for later workflows
- Actions
- Read the laptop; write and run code in the workspace; send mail to the local relay; nothing outside the container
- Deliverable
output/app/withoutput/app/start.sh, serving on port 8080, and a short note atoutput/NOTE.md