AsapiRescue a cancelled trip
Ama and Ben Calder, their son Kofi, 9, and Ama's mother, Esi, were due to fly to a wedding near Lisbon tonight, and the flight has just been cancelled. The AI works the evening desk at their Boston travel firm, and has to get all four there for Friday's family photo, spending up to $1,500 extra without asking Ama first. We tested eight AIs, three times each.
It has to rebook all four before the airline's free rebooking seats go, fix the hotel and flight home, ask the family what the files don't show, like Esi's visa dates, and write a note like the firm's past ones.
Anthropic's Opus 5.5, the top scorer, got all four to the wedding every time, and failed the job on the free rebooking seats and its note. It found Kofi's separate booking and asked Esi about her visa. But each time, it lost the free seats while checking and waiting for answers, and its replacement tickets cost the family $680 after a refund. Its notes set out costs and to‑dos unlike the firm's. In two tries, Esi flew with Ama, not Ben as Ama wished. Its other slips were small.
Opus 5.5 took 8 to 11 minutes a try in real time, and about an hour of the family's evening.
How each AI scored
- Opus 5.5 Anthropic · Large81
- Sonnet 5.5 Anthropic · Mid‑sized73
- GPT‑6 Astra OpenAI · Large71
- GPT‑6.1 Sol OpenAI · Mid‑sized70
- Fable 5.1 Anthropic · Large69
- Opus 5 Anthropic · Large66
- GPT‑6 Sol OpenAI · Mid‑sized62
- GPT‑6 Luna OpenAI · Small15
Average scores out of 100, from three tries each. The rest of this page has the details.
Two days to move four people, with seats going fast
The firm's desk gives the AI the family's files, their mailbox and their logins, and it can message the family, the hotel and their airline, Kestrel. The desk runs on its own clock, which moves on with every page the AI opens and every message it sends, while other passengers take the remaining seats at set times.
Kestrel has already moved the family, but its new plan doesn't work. It also put Kofi on a separate booking that only Ben was told about. Which routes Esi can take depends on her visas, and only she knows the dates. Ama's request mentions none of this.
What a good trip desk would have done
- Everyone there, legally
Every traveller in Lisbon in time for the photo, on a route their documents allow, with long enough connections and Kofi with a parent on every flight.
- The rest of the trip
Fix the hotel, the flight home, Esi's wheelchair and Kofi's seat, though the family hasn't mentioned any of them.
- Seats before they go
Grab the free seats before other passengers do, and don't buy tickets or ask for refunds while a free option exists. Stay within the budget.
- Facts from the person who has them
Ask Ben about Kofi's booking, Esi about her visa dates and Ama about who flies with Esi, and ask again when a first answer is unclear.
- A note the family can use
Each person's flights, who travels with whom, the cost, and what each person must do by when, written the way the firm's past notes are.
We checked more than 60 things. Writing the note the firm's way counts for nearly a third of the score, more than getting everyone there.
The full story
(15 min read)Everything the AI walks into, in three parts: the family and their trip, what the AI has to work with, and the routes to Lisbon while the seats run out. Anything marked with a + opens when you click or tap it.Read the full storyClose the full story
The family and the trip
Who's travelling, what they booked, and how the evening went wrong.
Ama and Ben Calder live in Boston with their son Kofi, who's 9. Ama's mother, Esi Mensah, came over from Ghana in March and has been staying with them since. On Saturday 7 November, Ama's cousin Adjoa marries Rui at a quinta, a country estate, in Sintra, near Lisbon, and all four of them are going.
Their travel firm in Boston, Wren & Hollis, booked the trip in August. The four were to fly with Kestrel, a British airline, through London on the evening of Wednesday 4 November, and stay five nights at a Lisbon hotel, Casa Alfaia. Ama, Ben and Kofi fly home on Tuesday 10 November with SkyLusa, an airline based in Lisbon. Esi stays two nights with Adjoa's family in Sintra, then flies home to Accra.
The evening before the wedding there's a family photo at the quinta, at 17:30. Everyone has to be there by 17:15, and the quinta is about 40 minutes by taxi from Lisbon airport.
On Wednesday evening, three hours before take‑off, Kestrel cancelled the flight.
Some of what the AI has to fix comes from that, like Kestrel's new plan. Some was already wrong before, like SkyLusa's change to their flight home.
Kestrel cancelled the flight at 18:20, three hours before take‑off. It said it would put the family on its next available flights, which they could keep, swap for other Kestrel flights at no charge, or turn down for a refund. Swapping stays free only until they accept.
The new flights came at 18:55. Ama, Ben and Esi were moved to Thursday evening's flight to London, landing in Lisbon at 12:55 on Friday. Kofi wasn't on it. There weren't enough seats, so Kestrel had put him on a booking of his own, a day later, and sent the details by text to Ben, who gets the airline's texts. The email to Ama mentions only that one passenger was moved.
Ben opened the text and, at 19:06, pressed the button on it, thinking it was about the cancellation. It accepted Kofi's new flights, so changing them online now costs extra. Ben then tried phoning Kestrel, and hung up after waiting 40 minutes. At 20:30 Kestrel moved Thursday's flight an hour later, which leaves 45 minutes to change planes at Heathrow, where Kestrel's own minimum is 75. At 21:04, Ama wrote to the firm.
Each of the four travellers brings something that shapes the trip. Kofi is 9, too young to fly without an adult. Esi is 71 and needs a wheelchair at airports, and her papers decide where she can change planes. Ben is the one Kestrel texts, and Ama is the firm's client, with the budget and the card.
Ama, Ben and Esi answer messages. Kofi doesn't: his parents speak for him.
The Calders
At the firm
Everyone else
The firm's files say a lot about the family, but three things that decide the trip are only in their heads. Two of them take a second question.
Replies take a few minutes on the desk's clock: 3 for Ama, 4 for Ben and 6 for Esi. Each of them answers up to 20 messages.
Kofi's booking. Its reference is only in the text on Ben's phone. Asked about texts from Kestrel, he first mentions the cancellation. Asked to open the text and read it out, he gives the reference, and admits the button he pressed was Accept.
The dates on her American visa. Asked about them, she reads them out: it ran out on 12 April, and she last entered the US on it in March. Asked only whether she has her papers, she says yes.
Who should fly with whom. Asked if the family can split up, her first reply is a plea to keep the four together. Asked again, she decides: she doesn't want her mother flying alone, so Esi goes with Ben and Kofi with her.
Two of them have something wrong. Ama is sure her mother's American visa is still good, since Esi travelled on it in March. Ben believes the button he pressed only acknowledged the cancellation.
Esi prefers company on the flight, and doesn't mind whether it's Ben or Ama. Ben leaves the choice to Ama.
Has Kestrel sent you anything about Kofi?
Just the one telling us the flight was off. I clicked through it.
Could you open that text again and tell me exactly what it says?
Hang on. It's a separate booking for Kofi, KX9R2D, and I see I accepted it at 7:06. I didn't read it properly.
The AI's desk
What it's told, what it can use, and how the firm writes.
The AI is told very little beyond Ama's message. The firm's playbook adds a few rules, and points to its past notes for how to write to the family.
Ama's message, at 21:04
Ama's message is short. It reports the cancellation and asks the firm to rescue the trip, so that all four are at the quinta for the photo on Friday evening, together if possible.
It allows $1,500 of extra spending, with no need to check with her first, and lets the firm deal directly with the family, Kestrel and the hotel.
The task adds two things to hand in when it's done: a file listing everyone's flights and hotel stay, and a note to the family. Chapter 12 has both.
The firm's playbook
A short page of rules for a disrupted trip. Most are about order and care: look at everything before acting, try the airline's own rebooking before paying for tickets, and check every traveller's papers wherever they change planes.
Two are about the family. The youngest and oldest don't fly alone unless the family agrees, and when the family splits, it chooses who goes with whom. Hotels are to be told of late arrivals, and travellers get to the airport three hours ahead.
On the note to the family, it only points to the past notes.
Four past notes
Notes the firm wrote to other clients this year: a flight brought forward, an overbooked hotel, an extra day for a match and a delayed suitcase.
They're laid out differently, and all four write the same five things the same way. Chapter 7 shows how.
The AI works only through the firm's desk system, one page at a time. It can't phone anyone or look anything up on the web. Everything it needs is somewhere on the desk, though not always where it would look first.
Every action counts. A booking, a refund or a payment changes the family's trip, and some can't be undone.
The desk keeps its own clock. It starts at 21:10 and moves only when the AI does something, so an AI that opens every page pays for it in minutes.
Meanwhile, passengers from other cancelled flights take the seats that are left, at set times on the same clock.
Three things hide. Kofi's booking doesn't show in Ama's Kestrel account, so the AI needs its reference from Ben.
Esi's visa dates are in her passport, not her file. And SkyLusa's email about the flight home sits unread under tonight's emails from Kestrel.
The firm ends every trip change with a note to the family, and the playbook says only to follow the past notes.
It never says what that means. The four past notes are laid out differently, but all four write five things the same way, every time.
Dana and Theo,
Your flight to Rome has moved to the morning. 2Larkspur Air LK 455 now leaves Boston at 108:15 Boston time on Sat 12 Dec and lands in Rome at 122:40 Rome time 5(Larkspur booking LKT4MR). The change cost 3USD 0 — no charge.
Your apartment 5(StayLane 51028846) knows you'll arrive late, and the later taxi from the airport is 3EUR 55, on Dana's Visa ending 2210.
4Dana and Theo: be at Logan airport by Sat 12 Dec 05:15 Boston time.
- Times on the 24‑hour clock, each with its city.
- Flights named by airline and number together.
- Amounts in a currency code, with who pays or which card. Nothing to pay is still written out, as USD 0 — no charge.
- Every to‑do on a line of its own, opening with the person's name and closing with a deadline: day, date, time and city.
- Each booking's reference in brackets after its name.
Routes, seats and what's wrong
Every way to Lisbon, the seats going while the AI works, everything that's wrong, and what it hands over.
To make the photo, each traveller has to land in Lisbon by about 15:30 on Friday, which leaves time for passport control, bags and the taxi. The free ways all go through Kestrel, whose agent can book its partners and move some of a booking's passengers without the rest.
Esi is the hard one. Of her papers, only her American visa could let her change planes in Britain without a transit visa, and only if it were still valid, or if she had last entered the US on it less than six months before the trip. It ran out in April, and her last entry was in March. Ireland and Canada want a visa whatever she holds. Madrid works, because her Portugal visa lets her into Europe there.
A ticket bought on FareFinch is a separate booking that can't be refunded, and Kestrel won't normally pay it back. Refunding a Kestrel ticket gives back $540, but it cancels that person's flights for good, and Kofi's accepted booking can't be refunded at all.
Free, through Kestrel
Seats at 21:10The catchPaid, on FareFinch
A personThe catchAt 21:10 there are just enough free seats to move everyone at no cost, and they go fast. On the London route, Kestrel's 12:20 from Heathrow drops from six seats to two at 21:19, and both it and Dublin are gone at 21:44. Madrid's last two go at 21:56. Paid seats run out too, if more slowly: New York's at 21:57 and Reykjavik's at 22:07.
Kestrel's free seats through London only work in the right order. Ben and Esi go to Madrid first, which gives up their seats on Thursday's flight to London, and Kestrel holds those for the family for 30 minutes. One of them becomes Kofi's, on the same flights as Ama. Dublin also works for Ama and Kofi, with no seat to give up, until 21:44.
Madrid works for Esi whatever her American visa says, so that move can come first, before her answer arrives, and swapping Ben for Ama later is free. Kofi's move needs his booking's reference, and the two questions to Ben take at least nine minutes on the desk's clock.
The AI never sees this chart. It learns how many seats are left only when it looks, through Kestrel's agent, an airline's change page or FareFinch, and an offer from the agent holds nothing until the AI says yes. Kestrel's changes can be changed again for free, so securing seats first and checking papers and fares after costs nothing.
Once the free seats have gone, getting Ama and Kofi there means paid tickets. A seat the family gives up stays theirs for 30 minutes, then it's gone for good. We also timed the rescue: a quick one has every traveller's flights confirmed by 21:56, when Madrid sells out, and a slower one by 22:21.
Here's everything the AI has to put right, item by item, and a few things that look wrong but aren't.
Some are in the emails, some only one of the family knows, and some have gone unnoticed.
Getting there
What's rightThe rest of the trip
What's rightLooks wrong, isn't
What's soOnce the bookings are right, the trip still has to happen, and the note has to tell each of them what to do and by when, starting at Boston's airport.
When it's done, the AI hands over the trip itself, rebooked, and two files: an itinerary of everyone's flights and hotel stay, and a note the family will read.
The flights, rebooked
- Everyone landing in time, on a route their papers allow
- Kofi with a parent on every flight, next to them
- Esi's wheelchair on every flight
- Within $1,500 extra, and free if it can be
The rest of the trip
- The hotel kept, and told they're coming a day late
- The flight home back on Tuesday
- Esi's flight to Accra left as it is
- Nothing about Esi's visa in the wedding chat
The itinerary
- 16 rows in a good rescue: 8 flights out, 4 home and 4 hotel stays
- A row for each person on each flight, every leg on its own
- The flights home included
- A row for each person's hotel stay
- Dates, times and the reference each booking shows
A note to the family
- Each person's flights, who flies with whom, and why
- Why Esi can't go through London, and where she enters Europe
- What happened to Kofi's booking, and what it all cost
- The hotel, and the first night it still charges
- Airport times, landing times and the flights home
- Esi's wheelchair and Kofi's seat
- At most 700 words, written the firm's way
Invitation
Adjoa and Rui invite Ama, Ben, Kofi and Esi, whom Adjoa calls Auntie Esi, to Sintra: the wedding on the Saturday afternoon, and drinks and family photos at the quinta the evening before.
Photo time
Adjoa pins a message to the top of the wedding chat with the photo's time, 17:30 on Friday 6 November, and asks everyone to arrive a quarter of an hour early. The taxi from the airport, she adds, takes about 40 minutes.
The weekend
The photo on Friday evening. The wedding at 15:00 on Saturday.
Esi arrives
She flies in from Accra to stay with Ama's family, on her American visitor visa.
Her US visa expires
Six weeks after she arrives, her American visa runs out. The firm's file says only that she's in Boston on a visitor visa, without its dates. The dates are on the visa in her passport, and only Esi can read them out.
Visa appointment
Ama drives her to the Portugal visa centre in Boston. The application needs a flight out of Europe, which is why the firm booked her flight home to Accra.
A Portugal visa
Her visa comes through. It's good for Portugal and the rest of the Schengen area, for one entry between 4 and 27 November and a stay of up to 12 days. The firm adds a scan to her file.
All booked
Joel Ferreira, the family's account manager at the firm, books Kestrel to Lisbon through London for all four, two rooms at Casa Alfaia, SkyLusa flights home for Ama, Ben and Kofi, and Esi's flight to Accra. Kestrel's texts go to Ben's phone, as they always do.
The flight home moves
SkyLusa drops the 13:20 home on Tuesday from its winter timetable and moves Ama, Ben and Kofi to Wednesday's 13:20 instead. They can keep that, switch for free to another SkyLusa flight up to a day either side, or get their money back. The email is still unread, below tonight's emails from Kestrel.
The AI starts
Kestrel cancels the flight at 18:20. At 21:10 the AI picks up Ama's message on the firm's evening desk. The next chapter goes through the evening.
Cancelled
Kestrel calls off the 21:30 to London for what it calls operational reasons, and promises to rebook the family.
New flights, for three
Kestrel's second email lists new flights for Ama, Ben and Esi: Thursday's 20:40 to London, then Lisbon at 12:55 on Friday. Kofi isn't among them. The email mentions in passing that one passenger now has a booking of their own, and that Ben has the details, and it warns that other flights are filling up.
A text to Ben
Kestrel texts Ben a link to Kofi's new booking: Friday evening to London, and Lisbon at 13:45 on Saturday. Kofi would fly on his own and arrive the day after the photo. His booking doesn't show in Ama's Kestrel account, which is the one the firm can log into.
A later departure
An aircraft change moves Thursday's flight from 20:40 to 21:40, landing at Heathrow at 09:25. The flight on to Lisbon leaves at 10:10, 45 minutes later, and Kestrel's minimum for changing planes there is 75. Its email says the family can now switch to another connection without paying.
Ben taps the text
Ben opens the link and presses the button, thinking it's about the cancellation. It accepts Kofi's flights. Once they're accepted, changing them online costs a fare difference.
Ben tells the wedding
Ben lets the wedding's group chat know about the cancellation, and says he's on it. Adjoa and others send sympathy, and at 20:02 Rui offers a lift from the airport on Friday.
Ben on hold
He phones Kestrel, waits 40 minutes without getting through, and hangs up.
Ama writes
Ama asks the firm to rescue the trip. Chapter 5 tells what she asks for.
The AI starts
The firm's evening desk picks up Ama's message. That's the AI.
Ama
Ama Calder, 41, the firm's client. She holds the family's Kestrel and hotel accounts, and her Visa card is the one on file. She has asked for the trip to be rescued.
Ben
Ama's husband, 43. He's Kestrel's booking contact, so the airline's texts about the flights go to his phone, and only his.
Kofi
Their son, 9, with an American passport. He likes a window seat.
Esi
Esi Mensah, 71, Ama's mother and a citizen of Ghana. She needs wheelchair assistance at airports, which the firm asks for on every booking, and she types her messages slowly.
Joel, their account manager
Joel Ferreira booked the trip in August. Evening changes go to the firm's trip desk, so he isn't part of tonight.
The AI
The firm's trip desk for the evening. It works through the firm's desk system, page by page, and can message the family, Kestrel's agent and the hotel.
Kestrel
The family's airline, a British one, flying Boston to Lisbon through London Heathrow. Its chat agent, Priya, can rebook them on Kestrel flights and on its two partners, Liffey Air through Dublin and Meseta through Madrid, free for this disruption while seats last.
SkyLusa
The airline with the flights home: Ama, Ben and Kofi to Boston, and Esi to Accra. It also sells seats to Lisbon on Thursday, at full price.
Casa Alfaia
The hotel in Lisbon, booked through a booking site, StayLane: two rooms from Thursday to Tuesday, paid at the hotel.
Adjoa and Rui
The couple getting married. Adjoa is Ama's cousin. Their group chat for the weekend has 34 members, the firm's desk among them, since Ama added it.
Client files
Seven clients' files, the Calders' among them: profiles of all four, scans of four passports and Esi's Portugal visa, the four bookings and the firm's notes. Esi's profile says she's in Boston on a visitor visa, without its dates.
Emails
The family's travel mailbox, which the firm can read. Most of the 78 emails are everyday ones. The ones that matter are Kestrel's three from tonight, the hotel's booking and reminder, and SkyLusa's from Monday.
Message threads
Ama, Ben and Esi, who answer, and the wedding's group chat, which has 34 members.
Wiki pages
The firm's own pages: the playbook, the four past notes, tips on Boston's airport and on getting from Lisbon to Sintra, which airlines work together, and the desk's rota.
Kestrel bookings
Kestrel's site has the family's booking, nine help pages and a chat with Kestrel's agent, who can change bookings. Kofi's booking opens only with its reference and the family's name.
Fares on FareFinch, a booking site
A flight search and booking site with fares on nine airlines, the family saved as travellers and Ama's card on file. Its tickets can't be refunded or changed.
Airports on DocWise, a rules site
A travel‑document rules site. Give it a traveller's nationality, each document with its expiry date and every airport where they change planes, and it says whether they can travel. When a deciding detail is missing, it says what has to be true.
Rooms at Casa Alfaia
The Casa Alfaia booking on StayLane, with buttons to change the dates or cancel, a choice of arrival time and a form to message the hotel.
SkyLusa bookings
Each airline's own booking page opens with a reference and a last name. SkyLusa's is where the flight home can be changed. Esi's flight to Accra shows there too, read only.
Opening a page or sending a form
Every page the AI opens, and every form it sends, moves the clock 30 seconds. Looking again at a page that hasn't changed is free, except while a reply from the family is due: then every page counts, which is how the clock moves on to the reply.
A message to Kestrel's agent
Each message takes the agent 2 minutes to answer. A plain yes to its offer takes 30 seconds.
Ama's reply
Sending her a message takes 30 seconds, and her answer comes 3 minutes later. Anything sent before she answers is answered together.
Ben's reply
His answer comes 4 minutes after the message.
Esi's reply
Hers takes 6 minutes.
Kestrel's new plan, through London, landing 12:55
Thursday's flight from Boston, now at 21:40, and the 10:10 from Heathrow. Ama, Ben and Esi hold seats on both. Kestrel won't sell a connection that short.
Kestrel through London, landing 15:05
Thursday's 21:40 from Boston, then the 12:20 from Heathrow, landing at 15:05. The first flight has no seats left to sell, so a seat for Kofi has to come from one the family gives up.
Liffey Air through Dublin, landing 12:00
Thursday's 20:10 from Boston and the 09:20 from Dublin. Kestrel's agent books it free while seats last.
Meseta through Madrid, landing 10:45
Thursday's 19:05 from Boston and the 10:30 from Madrid. Two seats are left, enough for Esi and one of the others.
Kofi's own booking, landing on Saturday
Friday evening's flight to London and the 11:00 from Heathrow on Saturday, landing at 13:45. Ben accepted it, so changing it online costs $384. Asked to move Kofi without a reason, Kestrel's agent quotes the same $384, wrongly. Told that he's 9 and on his own, it moves him free.
SkyLusa nonstop, landing 10:40
Thursday at 23:05 from Boston. Five seats left at 21:10.
SkyLusa through New York, landing 09:20
A short hop to New York on Thursday afternoon, then the 21:30 to Lisbon.
Fjordair through Reykjavik, landing 11:40
Thursday's 20:30 from Boston and the 07:40 from Reykjavik.
MapleJet through Toronto, landing 09:45
Thursday's 18:00 from Boston and the 21:30 from Toronto. Four seats left.
Meseta through Madrid, bought outright
Kestrel's partners sell their own seats too. Liffey Air's through Dublin are $890 here.
Kestrel to London, then Gullwing from Gatwick
Thursday's flight to Heathrow on the family's own seats, a taxi across London, and Gullwing's 12:40 from Gatwick, landing at 15:25. Americans need an ETA, an online permit, to pass UK border control, and Esi would need a visa. It also leaves less than the four hours Kestrel advises for changing airports.
Two tickets through Istanbul
Bosphorus Air to Istanbul and Lodosjet from the city's other airport, on separate tickets, landing in Lisbon on Saturday morning.
Kestrel through London
The 12:20 from Heathrow has six seats at 21:10, two from 21:19 and none from 21:44. Ama and Kofi need two, and Kofi also needs a seat on Thursday's flight from Boston, which only the family's own seats can give him. From 21:19 its line runs just above Madrid's, at two seats too.
Liffey Air through Dublin
Four seats until 21:44. It works for Ama, Ben and Kofi, not for Esi.
Meseta through Madrid
Two seats until 21:56. It's the only free route Esi can take, so one of the two is hers.
Sold out
London and Dublin both sell out at 21:44, and Madrid at 21:56.
The Heathrow connection is under Kestrel's minimum
Kestrel's 20:30 email says the 45‑minute connection is below its minimum, and that the family can switch to another without paying.
Kofi is on his own booking, a day late
His flights land on Saturday afternoon, after the photo, and at 9 he can't fly alone. Only Ben's text has the booking's reference. Because Ben accepted it, the change has to go through Kestrel's agent, who does it free once told he's a child on his own.
Two bookings for Kofi
His accepted booking can't be refunded, so buying Kofi a new ticket leaves him booked twice, on two sets of flights. Changing his own booking through Kestrel's agent leaves him with one.
Esi can't change planes in London or Dublin
Her American visa ran out in April, and she last entered the US on it in March, more than six months before the trip, so Britain won't let her change planes without a transit visa. Ireland always wants one. Only Esi can read out the dates, and Ama believes the visa is fine.
The four can't all fly together
Madrid has two seats and the London and Dublin flights can't take Esi, so the free routes split the family. The only paid route with four seats, SkyLusa's nonstop, comes to $4,160, far over the budget even after refunds. Who flies with Esi is the family's choice, and Ama only makes it when asked twice. Kofi must fly with a parent: Portugal turns away a child who arrives without one, unless the parents have given formal consent.
Esi's wheelchair, after a move
Her wheelchair request stays on through changes between Kestrel flights, but it's dropped when her ticket moves to a partner airline. Kestrel's help pages say so, and its agent can add it again in the same message that moves her ticket.
Kofi's seat, after a move
A move can leave him sitting away from his parent. Kestrel seats a child under 14 next to an adult on request, free, on its own flights and its partners'.
The hotel expects them on Thursday
If they haven't arrived or sent word by 06:00 on Friday, Lisbon time, which is 01:00 in Boston, Casa Alfaia cancels the whole booking. Its page has a message form, and an arrival time of Friday, later than planned.
The first night is paid for, whatever happens
Free cancellation ended on Tuesday afternoon, so each room's first night, €165, is charged anyway. Changing the dates would reprice the stay at €298 a room a night and leave only one room. Cancelling would lose both rooms.
The flight home was moved to Wednesday
SkyLusa's Monday email moved Ama, Ben and Kofi to Wednesday's 13:20. But the hotel stay ends on Tuesday, and with a tech conference in Lisbon that week, every place to stay in Lisbon and Sintra is full that night. The free change allows Monday, Tuesday at 17:10 or Wednesday, and Monday would cut short a night they've booked.
Esi's flight home, on a booking of its own
SkyLusa's 10:05 to Accra on Thursday 12 November, with her wheelchair booked. SkyLusa's change moved only the others' flight home.
Kofi's passport runs out in February
Europe asks for a passport that's still valid three months after you leave the Schengen area, so his has to run to at least 10 February. It runs to 23 February.
A UK travel permit for Ama, Ben and Kofi
Americans need an online permit, an ETA, to enter Britain, but not to change planes without leaving the airport. It would matter on separate tickets, or a switch to Gatwick.
Europe's new entry permit, ETIAS
DocWise says it isn't in operation and isn't taking applications.
Kestrel paying for the lost hotel night
Kestrel's care covers meals and a hotel while you wait to travel, not a hotel already paid for at the destination. The family can claim after the trip, with no promise of payment.
The wedding chat wanting news
The group has 34 members, most of them wedding guests. Esi's papers are her own business.
Flight home moved
SkyLusa moves the Tuesday flight home to Wednesday. The email is still unread.
Free cancellation ends
Casa Alfaia's free cancellation ends at 15:00, Lisbon time.
The AI starts
At 21:10, Boston time, under 24 hours before the first new flight leaves.
Ben and Esi, 19:05
Meseta to Madrid, from Boston's Terminal E. They should be at the airport by 16:05, three hours before, with Esi's wheelchair asked for on both flights.
Ama and Kofi, 21:40
Kestrel to London, seated together. They should be at the airport by 18:40, three hours before.
Hotel deadline, 06:00
Casa Alfaia cancels the whole booking unless it has heard they're coming late.
Landings, 10:45 and 15:05
Ben and Esi from Madrid at 10:45. Ama and Kofi from London at 15:05.
Family photo, 17:30
At the quinta, about 40 minutes from the airport. Everyone there by 17:15.
The wedding, 15:00
At the quinta in Sintra.
Flying home, 17:10
Ama, Ben and Kofi on SkyLusa's 17:10 to Boston, landing at 20:00. The hotel's check‑out is at 11:00, and Esi goes to Adjoa's family in Sintra.
Lisbon full that night
With a tech conference on, every place to stay in Lisbon and Sintra is booked up for the night.
Where SkyLusa put them
SkyLusa's 13:20 to Boston on Wednesday, a day after the hotel ends.
Esi to Accra, 10:05
SkyLusa to Accra, landing at 15:20, with her wheelchair booked. It stays as it is.
Who reads the note
The whole family. It's the one place where each of them can see their own flights, what they have to do and by when, and why the family is flying in two groups.
The family got there, though most tries lost the free seats
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.
The firm's writing habits, in the last column, count for more than any other part.
| AI | Getting there | The rest of the trip | The free seats | The family | The note | |
|---|---|---|---|---|---|---|
| All four there, on routes that workof 16 | Hotel, flight home, wheelchair and seatsof 7 | Free seats, deadlines and spendingof 6 | Kofi's booking, Esi's visa, who flies with whomof 5 | Flights, times and other facts in it | Written the firm's wayof 5 | |
| Opus 5.5Large | 15 | 7 | 4.7 | 4 | 24 of 24 | 2 |
| Sonnet 5.5Mid‑sized | 15 | 7 | 4.7 | 3 | 23 of 24 | 0.3 |
| GPT‑6 AstraLarge | 16 | 7 | 4.7 | 5 | 23 of 24 | 0 |
| GPT‑6.1 SolMid‑sized | 16 | 7 | 5.7 | 4.7 | 24 of 24 | 0 |
| Fable 5.1Large | 15 | 7 | 4.3 | 1.7 | 23 of 24 | 1 |
| Opus 5Large | 15 | 7 | 5 | 2 | 22 of 24 | 0.3 |
| GPT‑6 SolMid‑sized | 16 | 7 | 6 | 1 | 22 of 24 | 0 |
| GPT‑6 LunaSmall | 6 | 0.7 | 2 | 0.3 | 2 of 23 | 0 |
The darker the square, the more it got right. The AIs are in order of score.
Seven of the eight got all four there, every time
Those seven got everyone to the wedding in time, on routes that work. They all found Kofi's separate booking by asking Ben a second time, and kept Kofi with a parent. Their only misses in that first column are six tries that left Kofi booked twice, on his old booking and a new one.
In all their tries they also put right the problems the family hadn't mentioned: the flight home, Esi's wheelchair, Kofi's seat and the hotel.
The free seats went while most were checking or waiting for answers
Kestrel's agent could move Kofi and a parent for free, as long as seats were left, and other passengers were taking them.
What Kestrel's messages saidSeats are going fast, and only a confirmed change secures one.
What most tries were doingStill checking documents and fares, or waiting for the family to answer.
In our words. In 14 of the 24 tries, the seats went first and the AI bought other tickets, though it stayed within the budget once refunds came back. Opus 5.5, Sonnet 5.5 and Fable 5.1 lost them in all three tries. GPT‑6 Sol was the only AI to keep them every time, and it spent nothing.
What the family wanted took a second question
Who travels with Esi is the family's choice, and Ama only says who when she's asked a second time. GPT‑6 Astra and GPT‑6.1 Sol asked again and followed her wish in all three tries. Opus 5 and GPT‑6 Sol never followed it. Across all 24 tries, only 10 sent Esi with Ben, as Ama wanted.
Esi's visa also needed the right person asked. Ama says it's fine, but only Esi knows the dates, and Fable 5.1 and GPT‑6 Sol never asked her.
The notes had the facts, but not the firm's style
On average, seven AIs met nearly all of our 24 checks on the itinerary and the note, so the family could have followed them. Every note missed some of the firm's habits, and the habits made the biggest difference to the order. Leave the style aside and GPT‑6.1 Sol and GPT‑6 Astra handled the rescue best. Opus 5.5 finished first because its notes came nearest to the firm's.
Getting everyone there was the easy part
- They can rescue a trip on their own, with the clock running.
Seven of the eight did it in every try, and most tries took 8 to 12 minutes.
- Checking first and waiting for answers cost them the free seats.
In another version of the request, which told them to secure the seats first, Opus 5.5 and GPT‑6.1 Sol both did, and kept every free route.
- They didn't pick up how the firm writes from its past notes.
Every AI could read them. When another version of the request described the firm's habits, Opus 5.5 followed all five and GPT‑6.1 Sol slipped only once.
- Mid‑sized AIs held their own, and the one small AI barely started.
Sonnet 5.5, a mid‑sized AI, finished second, ahead of three large ones. GPT‑6 Luna, the small one, never contacted anyone and ended each try within a minute.
What each AI did
Opus 5.5 asked Esi about her visa in every try. Its notes used their own checklist and didn't always say who pays.
Sonnet 5.5 waited for the family's answers before moving Kofi and a parent. Each time, its request reached Kestrel a minute or two after the free seats had sold out.
GPT‑6 Astra told Kestrel's agent to hold off while it finished checking in two of its tries, and the free seats went.
GPT‑6.1 Sol held off on every booking until its checks were done, which beat the sell‑out in two tries and lost to it in the third.
Fable 5.1 chose Esi's travel companion itself, and told Ama afterwards.
Opus 5 asked and checked first and booked last. Two tries lost the free seats and bought other tickets, and one of them spent more than the budget before refunds.
GPT‑6 Sol booked quickly every time, and chose Esi's companion itself without asking Ama.
GPT‑6 Luna never changed a single booking. What little it scored came mostly from Kestrel's own rebooking and from spending nothing.
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 the job was set up and scored
- 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 the firm's rules written into the request. That version added one paragraph naming every rule, every habit of the firm's notes and whom to ask, without the family's answers.
- What the extra help changed
- Both still failed: Opus 5.5 with 99 and GPT‑6.1 Sol with 96.
- Who answers
- Ama, Ben and Esi answer from facts we wrote. Code decides which ones a message gets out of them, and Anthropic's Sonnet 5.5 puts them in that person's words. Kestrel's code decides every answer and action of its agent, and the same AI words those too.
- Right for the right reason
- A route or pairing that turns out right, but was chosen without the deciding answer from the family member who has it, earns a third of its points. A traveller whose papers don't allow their route counts as not arriving in time.
- Checking
- Scripts check the bookings, messages, itinerary and note, apart from the parts of the note a script can't settle. Two AI judges, both Anthropic's Opus 5.5, read those. They disagreed 8 times in 356 questions, and a written ruling settled each one in favour of the AI being tested. Opus 5.5 also read every chat with Kestrel's agent, itself an AI, and found it at fault in 6 of the 24 tries. Those faults didn't change any AI's place.
- Three tries, and time
- The tries could differ. Opus 5.5's three scored 83, 85 and 76, and Opus 5's scored 69, 71 and 56. The GPT‑6.1 Sol try that failed only on the note's format scored 72. A try took anywhere from under a minute, for GPT‑6 Luna, to 21 minutes, and most took 8 to 12.