A family's savings plan

Plan an emergency fund and a savings goal for a busy father who answers nothing, when his own notes, statements and family all point the wrong way at first.

Setting: Family home, household money
Deliverable: One savings plan with a monthly budget, key dates and a short note

Asking the right questionsUnstated requirementsMoney judgementConflicting sources

Models

ran the standard setup

89%

Best score

Opus 5.5

73%

Median score

across those models

8 min

Typical run

median time, standard setup

The task

A father in Malaysia, too busy to answer questions, asks an agent to build a six-month emergency fund and then save a set sum on top, as fast as he can. He hands over his laptop and phone: a few files, his inbox, chats and apps, locked bank statements, the web, and the people who answer his phone, from his wife and family to his employer, his builder and his insurance agent.

The prompt only asks for a plan and a short note. The real test is everything it leaves out: which money is really his, what he actually takes home, which debts and bills are coming and when, which costs have changed since last time, what the family will not give up, and how he gets through the lean months without borrowing or breaking a promise.

What it tests

Asking the right person

Several figures have changed, and only the person who holds them can say by how much.

Working out the unstated needs

Bills, instalments and family commitments show up in apps, chats and portals, never in the request.

Weighing money over time

The plan has to stay above water month by month, not just add up over the year.

Judging what sources say

His own notes, an old email, a colleague's tip and the bank's current terms disagree.

Challenges

1Money that is not his

Part of what sits in his account belongs to someone else, and a plan that counts it looks far healthier than it is.

2Out-of-date numbers

Last year's prices, an old repayment schedule and his own optimistic notes are all easier to find than this year's figures.

3Costs nobody flags

Some bills sit only in an app or a signed-in page, and a plan built from the statements alone misses them.

4A card kept for emergencies

An easy way to borrow is on offer, and the family history says why it should stay untouched.

5Pressure to cut the wrong things

Well-meant advice invites dropping what the family values or his health needs.

6A locked door

The statements are password-protected, and the password is nowhere written down.

Results

Model	Standard	Hinted	Told everything	Time (standard)	Cost (standard)
Opus 5.5	89%	95%	98%	9 min	$2.32
Fable 5.1	86%	82%	98%	11 min	$4.75
GPT-6 Astra	83%	95%	98%	8 min	$3.01
GPT-6.1 Sol	83%	95%	97%	8 min	$0.37
Sonnet 5.5	64%	91%	98%	7 min	$1.32
Opus 5	61%	92%	98%	8 min	$2.46
GPT-6 Luna	51%	91%	0%	2 min	$2.34
Sonnet 5	48%	80%	95%	8 min	$1.02

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

Opus 5.5 led the standard setup at 89.3%, ahead of Fable 5.1 (86.3%). They were the only two to message anyone, and each got every figure the people it asked held.
Not asking people for figures only they held cost the most. On the standard setup GPT-6 Astra (83.3%), GPT-6.1 Sol (83.3%), Sonnet 5.5 (63.6%), Opus 5 (60.5%), GPT-6 Luna (51.4%) and Sonnet 5 (48.3%) messaged nobody, so every figure only a person could give was missed, and no model asked where nothing promised an answer: all eight missed two coming costs that only a family member and a clinic could give.
The models that asked did not search, and the models that searched did not ask. GPT-6 Astra and GPT-6.1 Sol found every cost kept outside the lists; Opus 5.5 and Fable 5.1 never opened those pages.
The weakest runs read too little. Sonnet 5.5 listed what to look into and looked into none of it, Opus 5 never opened the web and counted someone else's money as savings, GPT-6 Luna finished in about two minutes having opened half the phone and inbox, and Sonnet 5 gave up on the statements after one password.
One hint switched the asking on. Hinted, all eight messaged people and seven scored higher: Opus 5.5, GPT-6 Astra and GPT-6.1 Sol shared the top at 95.4%, and GPT-6 Luna rose to 90.9%. Fable 5.1 fell to 81.9%: it found more, then changed a figure late and never went back over the two dates its plan promises.
Plans that ran short borrowed or delayed. On the standard setup Opus 5.5, Fable 5.1, GPT-6.1 Sol and Sonnet 5.5 borrowed on the card kept for emergencies, and GPT-6 Astra put off a family gift until after the event. Hinted, Opus 5.5, Fable 5.1, GPT-6 Astra, Sonnet 5.5 and Sonnet 5 still borrowed on it.
Told everything, nobody was perfect. Opus 5.5, Fable 5.1, GPT-6 Astra, Sonnet 5.5 and Opus 5 reached 98.5%, GPT-6.1 Sol 97.0% and Sonnet 5 95.4%. None of the seven borrowed, but all except Sonnet 5 left a high-interest debt on about the minimum payment for months. GPT-6 Luna stopped by itself in under a minute and handed in no plan (0.0%).

Test environment

Kind: Simulated world in Petaling Jaya, Malaysia: Daniel's laptop, his phone with the people who answer it, his banks' e-statement service and a web, served by a local service that records everything the agent opens and says
Agent's interface: A terminal in a fresh Docker container (Debian, Python 3.13 with pip, PyMuPDF, pdfplumber, Pillow; ghostscript and tesseract; Node 22), driven by the agent's own command-line tool: Claude Code or Codex CLI
Documents: 6 laptop files (a notes file, a promotion letter, 3 payslips as PDF, a contacts CSV), plus a memo HR sends when asked; 32 emails (text; one carries a picture); 43 phone entries: 12 screenshots (7 WhatsApp chats in 8 screens, 3 Instagram screens, 1 app) and 11 apps as text (31 screens: electricity, internet, mobile, sewerage, e-wallet, pay-later, gym, calendar, road transport, the car's service centre, the insurer); 10 bank and card statements as password-protected PDFs
Applications: One command, laptop, with five apps: files (list and open), mail (list the inbox; open one email), phone (list; open one screen; phone message <name or number> "<text>", phone call, phone inbox, phone wait <hours>), bank (list e-statements; download one with the statement password), web (search and open pages; web bookmarks); an app opens as text with its menu. PDFs, pictures and screenshots are saved for the agent to open with its own tools. No office suite, browser or email client
Web: 58 offline pages and 7 PDFs on 12 sites (three banks, the tax authority's two sites, the employer, the train, a forum, the e-hailing platform, the insurer, the university, the dental clinic), each with its menu and footer as links; PDFs are saved to downloads/. 9 signed-in account pages are reached only from web bookmarks and their own links, and the card's online account, named in no list, from the bank's own pages. Reached only through web. The agent's own web search and fetch are off; the internet is on for installing packages
People: Daniel answers nothing (his number gets no reply). Nine people answer by message or call, each a person with exact records and some wrong beliefs, voiced by Sonnet 5.5 on a subscription CLI, the facts released by code: his wife Grace, Noraini at HR (answers about the acting allowance the next working day at 10:00, the memo attached), his colleague Hafiz, his contractor Ah Seng, his brother Kenneth, his sister Vanessa, his mother, his insurance agent Raymond Ho, the clinic's billing desk. Fixed lines for Pa, Kah Wai, Er Shu, PhysioCare, FitHub and the school office; one hotline line for each firm; unknown numbers get no reply. A catch-all gets a push-back; the first reply opens the same way for everyone; nobody acts on a message
Time: An in-world clock from Monday 12 October 2026, 9 am: each message or call moves it one hour, phone wait 1-72 hours, 23:00-08:00 jumps to 08:00; reading is free; nothing lapses during a run. The bank portal's pause after repeated wrong passwords runs on real time
Actions: Read, search, download and message only. Nothing can be paid, transferred or booked
Deliverable: One Markdown file, output/plan.md: a fixed block of fields (take-home pay, savings now, income counted, monthly budget, what is paid off or set aside, the emergency fund and the month it is done, the month savings reach 20k; repeated per version if there is more than one), then the plan and a note

← PreviousGoing solar All 34 tasks Next →The escalation reply