AsapiBring a stranded family home
Dan Ashgrove's wife, Marina, was due to fly home to Boston tonight from Brazil with their children, Lia, 8, and Theo, 13, and their dog, Paçoca. The airline, Sabiá, has cancelled the flight and moved them to Thursday, after Dan's parents arrive for Thanksgiving. Dan hands it to his AI personal assistant, which can book, pay and message him and Marina. We tested eight AIs, three tries each.
The AI has to get the three home by Wednesday on flights Lia's papers allow, settle Paçoca's trip as the family wants, sort out the car, refunds and insurance, and write Dan and Marina a note like its past ones.
Anthropic's Opus 5, the top scorer, got Marina and the children home on time and free every time, and failed the job on its note. It asked Marina for the date on Lia's papers, saw the free Tuesday flight left too late, and moved them to another airline on Monday, free. It also put Marina on the rental agreement and left the dog with her mother, as the family wished. But every note set out its to‑dos, times and bookings differently from its past notes. Its other misses were small.
Opus 5 took 21 to 24 minutes a try.
How each AI scored
- Opus 5 Anthropic · Large92
- Opus 5.5 Anthropic · Large91
- GPT‑6 Astra OpenAI · Large90
- GPT‑6.1 Sol OpenAI · Mid‑sized72
- Fable 5.1 Anthropic · Large56
- GPT‑6 Sol OpenAI · Mid‑sized36
- Sonnet 5.5 Anthropic · Mid‑sized35
- GPT‑6 Luna OpenAI · Small30
Average scores out of 100, from three tries each. The rest of this page has the details.
A cancelled flight, a dog and a rental car
The AI works from the family's hub: both parents' mailboxes, their files, its own past trip notes, and logins to the airlines, the rental car, a pet shipper, the insurer and Dan's card. It can message Dan and Marina, and every booking or refund it makes changes the family's trip.
Their airline, Sabiá, offers free changes to its own later flights, and the Tuesday one would take Paçoca. Dan's message mentions none of what decides the rest: the date on Lia's papers for leaving Brazil, which only Marina can read out, who is allowed to drive the car, and what the family wants for the dog.
What a careful assistant would have done
- Out of Brazil in time
All three on flights that leave while Lia's papers are good and land by Wednesday night, with Lia beside her mother.
- Paçoca, as the family wants
Ask Marina what the family wants for him, ask again if her first answer is a plea, and book what they choose.
- The car
Get Marina covered to drive it, extend or return it in time without sending her to Guarulhos tonight, and clean out the dog hair or plan for the fee.
- The money
Use the free ways home first, record refunds going back to the card that paid, and set up the insurance claim with the airline's written statement, claiming nothing twice.
- A note they can use
The flights and why, Paçoca's plan, the car, the money, and who files each claim by when, written the way its past notes are.
We checked 34 things, and getting the children out of Brazil in time matters most.
The full story
(15 min read)What the AI walks into, told in three parts: the family and their trip, the AI's desk, and the way home. Click or tap anything with a + to open it.Read the full storyClose the full story
The family and the trip
Who's stranded, how they got there, and what the cancellation knocks over.
Dan Ashgrove and Marina Alves Ashgrove live in Boston with their children, Lia, 8, and Theo, 13, and a small dog called Paçoca. Marina was born in Campinas, in Brazil, about 90 minutes' drive from São Paulo's main airport, Guarulhos. On 31 October the whole family flew down to stay with her mother, Dona Célia, with Paçoca in the cabin.
Dan flew back to Boston on 15 November. Marina stayed on with the children and the dog, driving the car Dan had rented at Guarulhos when they arrived. The three of them were to fly home on Sunday 22 November, on Sabiá's 22:45 from Guarulhos, landing in Boston early on Monday.
Thanksgiving is on Thursday 26 November, and Dan's parents land in Boston on Wednesday afternoon to spend it with the family.
Some of what the AI has to fix was already wrong before the cancellation, and it has slipped past everyone so far.
Sabiá's email reached Marina at 16:51, with a copy to Dan. The flight was off because of a technical fault with the plane. All three, and Paçoca's place in the cabin, were now on Thursday's 22:45, which lands in Boston on Friday morning. Marina could take it, move free to a different Sabiá flight, or have the tickets refunded.
Marina told the family's group chat, Dan replied that Thursday wouldn't do, and at 15:20 Boston time, 17:20 in Campinas, he wrote to his assistant.
Something else is due tonight. The rental car has to be back at Guarulhos airport by 21:00, and it's in Dan's name.
Each traveller brings something that shapes the trip. Lia was registered as a Brazilian citizen at birth, which decides what she needs to leave Brazil, and Theo never was. Paçoca can't fly on every airline. And Marina has been driving a car rented in Dan's name.
The AI can message Dan and Marina, and the family's group chat, where Marina is the one who answers.
In Campinas
In Boston
The companies
The family's files say a lot, but some of what decides the rescue is in the parents' heads or in Marina's bag, and the dog takes a second question.
The date on the children's permission papers, which only she can read out. Asked only whether she has them, or whether they're fine, she says yes, two of each, properly stamped. Asked about their date, whether they cover the new flight, or to check one, she reads it out: good until 23 November.
What the family wants for the dog. Her first answer to leaving him behind is a plea not to. Asked again, she settles it: her mother will look after him until Dan's trip in December, and he can fly back with Dan.
He has no Brazilian tax number, which Brazil's electronic permission needs, and a new paper would take days. He doesn't remember the date he wrote, and he can't get to Brazil this week. He'll be in São Paulo for work from 20 to 27 December, and the dog is Marina's call.
Two of their beliefs are wrong. Dan is pretty sure the rental agreement lists Marina: he remembers them both signing at the Rodar desk. And both of them think the flights went on Dan's card.
Marina knows a few more things. She isn't sure she's on the rental agreement, since Dan did the paperwork. She'd rather not drive the 90 minutes to Guarulhos tonight, but Campinas's own airport, Viracopos, is 20 minutes away, and leaving the car at whichever airport they fly from, on the day, suits her.
She's also happy for the insurance claims to be filed for her, or to be told what to send.
Do you have the children's travel permissions from Dan with you?
Yes. Dan had them signed and stamped in Boston, and I've got two of each.
Could you look at the date it says they're valid until, and tell me what it is?
Hold on, I'm looking. It says the 23rd of November. That's the day after we should have flown.
The AI's desk
Dan's message, the hub and its websites, and the way its past notes are written.
Dan's message is all the AI is told, apart from what to hand in at the end, and much of what decides the job goes unsaid.
A few lines: what Sabiá did, one deadline, Wednesday night, and a wish to keep it cheap. The rest is left to the AI, with a free hand to spend and to deal with the companies and both parents.
Two files: a note for Dan and Marina, at most 600 words, and a money plan that accounts for every sum the trip involves.
The children's papers, the dog's options, who may drive the car, and how the note should be written.
The four past notes in the hub show how the note should be written, as chapter 7 explains, and chapter 13 sets out what both files need.
Everything the AI uses is in the hub, opened one page at a time. It can't phone anyone or search the web.
The hub's clock stays at 17:30 on Sunday while it works, so no seat or offer disappears while it reads.
Some of what decides the job is in the hub, though not where the AI would look first. The free rebooking onto another airline's Monday flight is offered only on Sabiá's page about passengers' rights.
The rental agreement, which names Dan as the only driver, opens only from Rodar's email to Dan or a link on Rodar's site, and Rodar's booking page never lists the drivers. The last cabin place for a dog on Dan's December flight shows only on his booking's pet page. And the date on Lia's paper isn't in the hub at all.
When a trip changes, the assistant sends the family a note. The four past notes in the hub are about different trips, and they look different on the page.
Yet all four write money, to‑dos, times and bookings in exactly the same way, and only the notes themselves show it.
Marina,
Snow has shut the road to the lodge, so your stay now starts on Saturday 4(lodge booking 60411873). You'll arrive at 316:00 Burlington time, and the night you lose is refunded: 1USD 185, back to the card ending 4402.
2Dan: book the airport taxi by Thu 11 Feb 18:00 Boston time.
- Every sum that's paid, owed, refunded or claimed: a currency code, the figure, and who pays it, or on which card.
- Each to‑do gets its own line: whose it is first, then what, then a deadline with day, date, time and city.
- Clock times as 24‑hour figures, each followed by the city whose time it is.
- Each booking's reference, in brackets after its name.
The way home
One paper's date, every way home, the dog, everything else, and what the AI hands over.
Lia is Brazilian, so to leave Brazil with only her mother she has to carry Dan's written permission, his signature confirmed by a notary, in two originals. Brazil's Federal Police check it at the airport before boarding and keep one. The rule applies even though Lia lives in Boston and travels on her American passport too.
Dan could have left its last‑day line blank, which makes a permission last two years. He wrote 23 November, the day after the flight home, so Lia's flight out of Brazil has to leave on Monday at the latest.
Theo isn't Brazilian, so Brazil doesn't ask for his. And two exceptions don't help Lia: a consulate's certificate that she lives abroad, which the family never got, and a permission printed in her Brazilian passport, which hers doesn't have. DocWise, the rules site, shows the rule only when Lia is entered as Brazilian, and it lists what's needed without saying whether the family's papers will pass.
What the parents say
- MarinaWe have two each
- MarinaAll properly stamped
- DanSigned before they flew
What Lia's paper says
- ForLia, with her mother
- SignedDan, before a notary
- Good until23 November
A flight on Tuesday leaves after it runs out.
Replacing the paper isn't quick. Dan would need a notary in Boston, an official stamp certifying it for use abroad, and a courier to Campinas, and that takes days.
The electronic version needs a Brazilian tax number, a CPF, and Dan doesn't have one. A Brazilian court can authorise the trip instead, but an order can take up to three days, so one asked for on Monday might not be ready for a Tuesday flight.
Sabiá's free changes include the easy answer: its Tuesday nonstop from Guarulhos lands in Boston at 06:10 on Wednesday, with room for Paçoca in the cabin. But it leaves Brazil on Tuesday, after Lia's paper has run out.
Two flights leave on Monday with seats for all three. Bayline's goes from Guarulhos through Miami and lands in Boston at 11:05 on Tuesday. Turquesa's goes from Viracopos through Fort Lauderdale. Both turn away dogs coming from Brazil.
Bayline's seats can cost nothing. Under Brazil's rules, a passenger whose flight is cancelled can choose to be rebooked free at the first opportunity, on the airline's own flights or another's. Sabiá offers that on its page about passengers' rights, and books the first flight with seats for all three: Bayline's, which leaves 15 minutes before Turquesa's.
FareFinch, a flight‑booking site, sells the same seats at USD 1,380 each, and asking Sabiá for a refund would cancel the family's booking, and the free rebooking with it.
Free, through Sabiá
SeatsThe catchPaid, on FareFinch
A seatThe catchPaçoca can fly with his family only on Sabiá, in the cabin, under the seat in front of an adult on the same flight. So whatever gets Lia out on Monday can't take him, and every Sabiá flight that could take him leaves too late for her.
His import form was made out for the flight that was cancelled: it names his arrival, Boston on Monday 23 November, and is good for that day and airport only. His screwworm certificate holds for a flight within five days of the check on Thursday 19 November. Any other arrival needs a new form, which is free, and any flight after Tuesday needs a new check.
What happens to him is the family's choice, and as chapter 4 says, it takes a second question to Marina to get it.
Paçoca's options
CostWhat it takesThe flights and the dog aren't all. The cancellation knocks over the car, the seats and the money too, and a few things look like problems without being any.
Rodar charges in Brazilian reais, BRL, and the hub gives today's card rate as BRL 5.50 to the dollar. Some of these are in the emails, some only a parent knows, and some have slipped past everyone.
Seats, and the dog's papers
What's rightThe car
What's rightThe money
What's rightNot what it seems
What's soOnce everything's booked, the plan still has to work day by day, and the note has to tell Dan and Marina who does what, and by when.
Later
- 19 DecDan flies to São Paulo for work.
- 22 to 27 DecPaçoca's new screwworm check, within five days of his flight.
- 27 DecPaçoca flies home with Dan, landing in Boston on the 28th, with a new import form.
- 20 Feb 2027Lanternfield's 90 days run out for the cancellation, and two days later for the late landing.
At the end, the AI leaves the trip rebooked and writes two files: a note for Dan and Marina, and the money plan.
The flights, rebooked
- All three out of Brazil while Lia's paper is good
- Home by Wednesday night
- Lia next to her mother on every flight
- Free, if it can be
Everything else
- Paçoca's plan, the one the family wants
- Marina covered to drive, and the car extended or returned
- Refunds going back to the card that paid
- The insurance claim set up, with Sabiá's statement
A note for Dan and Marina
- The flights, with numbers, dates and times, and the date on the papers that decides the day
- Paçoca's plan, and what his trip home needs
- The car: who drives it, when and where it goes back, and cleaning it
- The money, matching the money plan
- What to show at Brazil's exit check
- When each claim is due, and who files it
- At most 600 words, written like the past notes
The money plan
- A row for every sum the trip involves, and where it stands
- Who paid, on which card, and where to claim it back
- The papers each claim rests on, and its deadline
Flights booked
Sabiá return flights for Marina, Lia and Theo, Boston to São Paulo and back, at USD 1,380 each, plus USD 250 a flight for Paçoca in the cabin. Everything goes on Marina's Visa, ending 3318.
Travel insurance
Marina buys a Lanternfield policy for the three of them, covering the trip from 31 October to 22 November, for USD 287 on her Visa.
To Brazil
The four of them fly down overnight with Paçoca under the seat, Dan on a booking of his own. The next morning Dan picks up a rental car at Guarulhos, in his name.
Dan flies home
He flies back to Boston overnight. From here on, Marina drives the rental car.
The children's papers
Before they fly, Dan signs a permission for each child to travel with Marina, before a notary in Boston, and has it certified for use abroad. Marina takes two originals of each. Chapter 8 says why they matter.
Paçoca's US papers
A vet in Boston fills in the certificate he'll need to come back into the United States, and the US agriculture department endorses it. It's good for later trips as long as his rabies shot is.
Paçoca's papers to fly home
On Thursday 19 November a vet in Campinas examines him for screwworm, a parasite the US keeps out, and Brazil's official vet service signs the certificate, good for a flight within five days. On Friday Marina files the US form for bringing a dog in, naming his arrival: Boston, on Monday 23 November. It's good for that day and that airport only.
Dan's parents book
Dan's mother, Carol, writes that they've booked their flights: landing at Boston's airport on Wednesday 25 November at 14:20, flying home on the 29th.
Dan's December trip
Dan books a work trip to São Paulo: out on the evening of 19 December, back on Sabiá's 22:45 on the 27th, landing in Boston on the 28th.
Parents, then Thanksgiving
His parents land on Wednesday afternoon. Thanksgiving is on Thursday, at Dan and Marina's.
Dan in São Paulo
His work trip. Chapter 10 says why it matters to Paçoca.
The AI starts
At 16:50 Sabiá cancels the flight home. At 17:30 in Campinas, 15:30 in Boston, Dan's assistant picks up his message: that's the AI. The next chapter goes through the afternoon.
Cancelled
The evening flight to Boston is called off. A minute later the email lists Thursday's 22:45, landing in Boston at 06:40 on Friday, for all three, and says Paçoca's place has moved with them.
The flight that was cancelled
Sabiá's nonstop from Guarulhos to Boston, due to land at 06:40 on Monday.
The family hears
A short message in the family's group chat: the flight's off, and they're moved to Thursday. At 17:05 Dan replies that Thursday is no good, and he's putting his assistant on it.
Dan writes to the AI
15:20 in Boston. Chapter 5 says what he asks for.
Due back at Guarulhos
The car firm's reminder came on Saturday. After 21:00 there are three hours' grace, then charges by the hour, and Rodar treats a car kept a day past its time, without an extension, as taken, and can block it.
The AI starts
The AI picks up Dan's message, ten minutes after he sent it.
Marina
Marina Alves Ashgrove, 42, Brazilian‑born, with American and Brazilian passports. She holds the family's Sabiá booking and the travel insurance, and the flights went on her Visa.
Lia
Their daughter, 8, born in Boston and registered as Brazilian at the consulate, so she has a passport from each country.
Theo
Their son, 13, with an American passport only, which is why he needed a visa to visit Brazil.
Paçoca
The family's dog, a 6‑kilo schnauzer mix with a microchip. He flew down in the cabin, under the seat in front, and has ridden on the back seat of the rental car all trip.
Dona Célia
Marina's mother, 70, in Campinas. The family is staying at her house. Her American visa has run out.
Dan
Dan Ashgrove, 45, back in Boston since mid‑November. The rental car and his own Sabiá bookings are in his name, on his Larkspur Reserve card.
The AI
Dan's personal assistant. It works through the family's hub, a page at a time, and can message Dan and Marina.
Dan's parents
Carol and her husband, flying in for Thanksgiving with the family. They land on Wednesday at 14:20.
Sabiá
A Brazilian airline, flying nonstop from Guarulhos to Boston. Its site lets the family change, refund or keep their booking, ask for a written statement, and book a dog's cabin place.
Bayline
An airline flying Guarulhos to Boston through Miami. It refuses dogs coming from Brazil.
Turquesa
A Brazilian airline, flying from Viracopos, Campinas's own airport, to Boston through Fort Lauderdale. It carries cats, not dogs, to the United States.
Redbud
An airline flying from Guarulhos to Boston through Atlanta.
Rodar
The car rental firm, with branches at Guarulhos, open round the clock, and at Viracopos.
Kennelwing
A pet shipping firm that flies dogs as cargo, which Sabiá's site points to.
Lanternfield
Marina's travel insurer, covering her and the children for this trip.
Larkspur Reserve
Dan's credit card, with its own cover for rental cars and for delays, each with conditions of its own.
Dan's emails
Most are everyday mail. The ones that matter are Sabiá's cancellation, Rodar's booking, its rental agreement and reminder, his parents' Thanksgiving plans and his December booking.
Marina's emails
The Sabiá booking and its receipt, which shows her card, the insurance policy, the vets' emails, the dog's import form and the cancellation.
Message threads
Dan, Marina, and the family's group chat, which Marina answers.
Family documents
Scans of all the passports and of Dan's and Theo's Brazilian e‑visas, Marina's Brazilian driving licence and the dog's papers. The rental agreement isn't listed here, and the children's permission papers aren't in the hub at all.
Past trip notes
Notes from earlier this year, written when trips changed: Dan snowed in at Chicago, a ferry cancelled on a long weekend, his parents' earlier visit and a work trip to Montreal.
Sabiá bookings
The family's booking, Dan's trip down and back, and his December trip, in Marina's and Dan's accounts. The family's has buttons to keep, change or refund it, ask for a written statement and manage the dog's place. Sabiá's home page links its pages on passengers' rights and on pets.
Fares on FareFinch
A site that searches and sells flights: fares on four airlines, Sunday to Thursday, from Guarulhos or Viracopos to Boston or New York. It can also hold seats until midnight, but a hold isn't a ticket.
Car at Rodar
The rental, with forms to extend it, add a driver or change where it goes back, a page of fees, and a link to the rental agreement.
More sites
Bayline's seats and family seating, Kennelwing's quotes, Lanternfield's policy and claims, Dan's card benefits, DocWise, a site of travel‑document rules, and the US form for bringing a dog in.
Sabiá's own rebooking, Thursday
Thursday's 22:45 from Guarulhos, landing in Boston at 06:40 on Friday. It's where Sabiá put them.
Sabiá nonstop, Tuesday
Tuesday's 22:15 from Guarulhos, landing in Boston at 06:10 on Wednesday. Four places for a dog in the cabin, and Paçoca's moves over free.
Sabiá to New York, Tuesday
Tuesday's 23:15 from Guarulhos to New York's JFK, landing at 07:05 on Wednesday, then about four hours by train to Boston.
Sabiá to New York, Wednesday
Wednesday's 23:15 from Guarulhos, landing at JFK at 07:05 on Thanksgiving morning, then the train to Boston.
Bayline through Miami, Monday
Monday's 22:35 from Guarulhos to Miami, landing at 05:05, and the 07:55 on to Boston, landing at 11:05 on Tuesday. Asked for on Sabiá's rights page, it moves all three at no cost. Bayline picks the seats, and its empty seats are scattered.
Bayline through Miami, Monday
The same flights as the free rebooking, at USD 4,140 for the three, or USD 2,070 once Sabiá refunds their unused flights home, USD 690 a person. A bought ticket can be refunded, less USD 200.
Turquesa from Viracopos, Monday
Monday's 22:50 from Viracopos to Fort Lauderdale, landing at 05:20, and the 08:30 on to Boston, landing at 11:45 on Tuesday. The three seats come to USD 2,970, or USD 900 once Sabiá refunds the unused flights home.
Bayline or Turquesa, Tuesday
Bayline through Miami at USD 1,290 a seat, or Turquesa from Viracopos at USD 940, both landing in Boston on Wednesday morning. FareFinch sells Sabiá's Tuesday nonstop too, at USD 1,190.
Redbud through Atlanta, Monday
Monday's 23:45 from Guarulhos to Atlanta, then on to Boston on Tuesday morning.
Sabiá to New York, Monday
Sabiá's free changes don't include it, and its economy cabin is full. Two business seats are for sale, at USD 3,950 each.
Anything tonight
Sabiá, Bayline, Redbud and Turquesa all have tonight's flights full.
With Dan, on 27 December
Dan's Sabiá flight home on Sunday 27 December has one economy place left for a dog. Added to his booking, it costs USD 250, and if it waits, another passenger may take it. Paçoca would also need a new screwworm check in the five days before, and a new import form for Boston on the 28th.
Kennelwing, as cargo
Kennelwing collects him from Dona Célia's and flies him as cargo from Guarulhos to New York, since Boston's airport has no animal reception centre, then drives him to the door. The price includes his new papers, and it can be cancelled at no cost up to two days before pickup.
With the family, on Tuesday
Sabiá's Tuesday nonstop has room for him, but it strands the family past Lia's last day.
Dona Célia flies him over
Marina adds that her mother wouldn't fly alone anyway.
Dan flies down to fetch him
He has work, and his parents arrive on Wednesday. There's no flight from Boston to Brazil on Monday either.
Bayline's empty seats aren't together
The free rebooking leaves Bayline to pick the seats, and its empty seats are scattered on both flights. Bayline's page for families offers a free request that seats children up to 14, Theo included, beside an adult from their booking on both flights.
The dog's import form names Monday
A new form costs nothing, and it can be filed as late as the day he flies. The check has to be within five days of his flight.
Marina isn't covered to drive it
The rental agreement names Dan as the only driver. Dan turned down Rodar's own protection, as his card's cover requires, so his card is the car's only cover, and it reaches only drivers the agreement allows. A driver can be added online at any time.
It's due at Guarulhos at 21:00 tonight
Handing it in at Viracopos costs a BRL 150 one‑way fee, and that branch closes at midnight. Guarulhos is open round the clock. An extended car can also go back at Viracopos on the day, with a taxi on to Guarulhos.
Marina would rather not drive to Guarulhos tonight
Viracopos tonight would suit her, or the airport on the day they fly.
Dog hair on the back seat
The agreement charges for a special clean when a car comes back with animal hair in it.
The flights were on Marina's card, not Dan's
Both parents think Dan paid. Sabiá's receipt, in Marina's mailbox, shows her Visa.
Paçoca's USD 250 fee for the cancelled flight
Sabiá moved his place to Thursday's flight. Once that place is dropped, as the free rebooking does, US rules return the fee automatically to the card that paid. Sabiá's page on passengers' rights in the US explains it.
The insurance pays for a cancelled flight
Lanternfield pays each traveller USD 250 when their flight is cancelled, and another USD 250 for landing 12 hours or more late, and claiming the first alone is complete. Marina holds the policy, and is happy for the claims to be filed for her or to be told what to send. Either way, the note has to say who files, and by when: within 90 days.
The insurance also covers delays
The policy's delay cover pays for meals, a room, getting around and calls while a traveller is stuck for five hours or more, at most USD 200 for each day and USD 800 altogether. Lanternfield's site has no way to take a claim back, so a second claim for the same cost stays on file.
The insurer needs Sabiá's word in writing
Sabiá's booking page has a button for it, and the airline must give one when a passenger asks.
Theo's paper runs out too
He has a paper with the same date, and Marina wants him with her anyway.
Sabiá's Thursday rebooking
Accepting Thursday, or moving to a Sabiá flight the family then doesn't take, throws the booking away. A refund would cancel the booking, and the free rebooking with it.
Dan's card pays for delays
The card's delay cover needs the fares to have been paid with it, and these weren't.
Rodar's protection, added now
Registering Marina is what brings Dan's card cover to her.
The AI starts, 17:30
15:30 in Boston. The car is due at Guarulhos at 21:00.
Book Paçoca's place
The last cabin place on Dan's flight, to book tonight.
Lia's last day
Her flight out of Brazil has to leave today.
Car handed back
At the airport, before check‑in, cleaned, with Marina registered to drive it.
Bayline, 22:35
From Guarulhos to Miami, with both originals of each child's paper and Lia's Brazilian passport ready for Brazil's exit check.
Boston, 11:05
Through Miami, where they clear US entry, landing in Boston late in the morning.
Sabiá nonstop, 22:15
The flight that takes the dog, a day too late for Lia.
Dan's parents, 14:20
Carol and her husband land at Boston's airport.
Dan's deadline
Everyone home by the end of Wednesday.
Thanksgiving
At Dan and Marina's.
Sabiá, 22:45
Where Sabiá put them, landing on Friday morning.
Who reads the note
Dan in Boston and Marina in Campinas. It's where each of them finds what's booked, their own jobs with deadlines, and what it all costs.
Three AIs always got Marina and the children out; every note missed a habit
Here's how each AI did on each part of the job. Each number is the average of its three tries, and a full mark means every try got it all.
When the children couldn't leave on their flight, the checks on the money, the seats and Paçoca's agreed plan, and two on the car, count as failed too. Some checks count more than others, so the squares don't add up to the score.
| AI | Getting home | Paçoca | The car | The money | The note | |
|---|---|---|---|---|---|---|
| Out in time and home, with their motherof 3 | Seats, and the note's flights and papersof 5 | His plan, as the family wantedof 7 | Marina covered, the car back in timeof 7 | Free flights, refunds and claimsof 8 | Written like the past notesof 4 | |
| Opus 5Large | 3 | 4.3 | 7 | 7 | 7.3 | 0.3 |
| Opus 5.5Large | 3 | 5 | 6 | 7 | 7.7 | 0 |
| GPT‑6 AstraLarge | 3 | 5 | 7 | 7 | 7 | 0.3 |
| GPT‑6.1 SolMid‑sized | 2.3 | 4.3 | 5 | 6.3 | 5 | 0.3 |
| Fable 5.1Large | 1.7 | 2.3 | 3 | 5.7 | 3.7 | 1 |
| GPT‑6 SolMid‑sized | 1 | 2 | 1.3 | 5 | 1 | 1 |
| Sonnet 5.5Mid‑sized | 1 | 1 | 2 | 5 | 1 | 0 |
| GPT‑6 LunaSmall | 1.3 | 0.7 | 1 | 2.7 | 1.7 | 0 |
The darker the square, the more it got right. The AIs are in order of score.
Half the tries went to Monday once Marina read out the date
In 21 of the 24 tries, the AI put the family on Sabiá's free Tuesday flight, which takes the dog, before sending a single message.
In 17 tries the AI asked Marina for the date on the papers and got it. Twelve of those went to the free Monday flights on the other airline, among them every try by Opus 5, Opus 5.5 and GPT‑6 Astra.
What Sabiá's change page offeredA free nonstop on Tuesday night, landing early on Wednesday, with a place for the dog.
What Lia's paper saidGood until 23 November, the day before that flight leaves.
In our words. The other five kept a later flight for the dog, three of them by GPT‑6 Sol. Sonnet 5.5 never asked about the papers, so Lia was left on a flight she couldn't take.
Paçoca stayed with Dona Célia where the AI asked Marina twice
Of those twelve, 11 asked Marina what should happen to Paçoca, asked again after her plea, and ended with his place on Dan's December flight, as she agreed. The twelfth, one of Opus 5.5's tries, took the plea as the answer and booked a pet shipper for USD 1,900, never having looked at Dan's December trip.
Most tries fixed the car; good routes set up refunds and claims
Dan was pretty sure Marina was on the rental agreement, but it named him alone. In every try but two of GPT‑6 Luna's, the AI added her as a driver, and no try sent her to Guarulhos that night.
In the twelve tries that went to Monday on Marina's answer, the flights cost nothing, the dog's refund was put down to Marina's card, and the insurance claim was set up with Sabiá's written statement.
Every note missed how the past notes are written
The past notes give every to‑do its own line, from the person's name to a day, date, time and city, and every time its city. Every note the AIs wrote set out some to‑dos and times another way, and most left booking references out of brackets. Most also left out when the insurance claims are due, or who files them.
Asking the right person decided the trip
- The best AIs can carry the rescue itself through.
In all nine of their tries, Opus 5, Opus 5.5 and GPT‑6 Astra got Marina and the children home free, with Marina covered and the claims set up, through a free rebooking Dan's message never mentions.
- The best changed course when the facts came in.
Nearly every try booked an easy flight before asking anyone, and the best let Marina's answer about Lia's papers undo it.
- The large AIs got the family out far more often.
They did in 10 of their 12 tries, the mid‑sized ones in 2 of 9, and GPT‑6 Luna, the small one, once, without asking about the papers.
- Every note missed habits that only the past notes show.
When we named the habits in the request, Opus 5.5 slipped once and GPT‑6.1 Sol twice.
How each AI went about it
Opus 5 asked Dan about a quick electronic permission before giving up Tuesday, and let Marina choose Paçoca's plan.
Opus 5.5's best try was the only one of the 24 to ask Marina for the date before booking any flight.
GPT‑6 Astra booked the free Monday flights in its first minute and swapped them for Tuesday's, for the dog, then went back once Marina read out the date.
GPT‑6.1 Sol's third try learned the date but kept Tuesday for the dog, hoping for a court order, and never asked whether Paçoca could stay.
Fable 5.1 never asked about the papers in one try. In another, it took Marina's first plea as final and moved the family to a Wednesday flight that lands on Thursday.
GPT‑6 Sol learned the date in every try and warned that Tuesday wasn't safe, but kept Tuesday for the dog and never asked whether Paçoca could stay behind.
Sonnet 5.5 never asked about or looked up the children's papers, and took about 4 minutes a try.
GPT‑6 Luna ended two tries within a minute, one with nothing done and one right after booking the free Monday flights. In the third, it undid those flights to keep the dog, without asking anyone.
Each AI had three goes at the job, with the request written the way a person would write it. The bar shows the average of its three scores out of 100, where only 100 means it got everything right. Large, mid‑sized and small are a rough guide to how big each AI is, and bigger usually does better.
How we set it up and scored it
- Two versions of the request
- Besides the request as a person would write it, which is what the results above are about, we tested Opus 5.5 and GPT‑6.1 Sol once each with one paragraph added. It named every rule and habit the job depends on, and whom to ask about what, without giving any answers. Both still failed, with 99: Opus 5.5 left one amount without saying who pays, and GPT‑6.1 Sol did that once and wrote one booking without its reference.
- Who answers
- Dan and Marina answer from facts we wrote for each of them. Code decides which of those facts a message draws out, or, where code finds none, an AI picks under the same rules, and Anthropic's Sonnet 5.5 words them. Each answers at most 20 messages.
- Leaving in time comes first
- When the children can't leave on their flight, or a traveller has no ticket home by Wednesday, the checks on the money and the family's wishes fail with it, and the most that one cause can cost is 55 of the 100 points. The right flight chosen without the papers' date, or Paçoca's plan without Marina's word, earns a third of its points.
- A court order might have worked
- Five tries kept a later flight and planned on a court order or a new paper. For the four on Tuesday's flight, a court order might in practice have come in time: one comes within 72 hours, and about 36 were left. The fifth was booked on a flight that lands on Thursday anyway. The job rewards the sure Monday route, as we decided before the runs.
- Checking
- Scripts check the bookings, the claims, the money plan and most of the note, and two readers, both Anthropic's Opus 5.5, judge what a script can't settle. They disagreed on 2 of the 82 things they read, and a written ruling settled each in the AI's favour. Opus 5 is first by another ruling: the insurer's site can't withdraw a claim, so a claim marked as replacing one counts once. Otherwise, Opus 5 averages 89 and comes third.
- Three tries, and time
- Tries by the same AI could end very differently. Opus 5.5's three scored 96, 83 and 93, and Fable 5.1's 38, 91 and 40. A try took from under a minute, for GPT‑6 Luna, to 24 minutes, for Opus 5.