Return on tokens: our answers, with real numbers

Her question, and her own example

Her worked example is worth restating exactly, because it makes money:

• 1,000 requests
• $0.60 of model and tool cost per request, paid on all 1,000
• 10% of requests end in a booking that sticks
• $30 earned per booking

That is $3,000 in and $600 out. Her real argument is not that the math fails. It is that nobody is measuring it, and that the cost of all the looking lands somewhere. She quotes a useful rule: “Agents should do the thinking; code should do the doing.” And she names the number that matters: “Cost per completed trip, including the failed attempts and the fixes, is a strategy number.”

So we measured cost per completed trip, failed attempts included.

What a trip costs us

Just Travel is a personal AI travel agent. You plan in conversation, compare live flights, hotels and places side by side, and make every decision yourself. Every model call a trip makes is recorded with its token counts, so we can price each one.

We ran 12 benchmark trips covering the range people actually plan: a weekend in Lisbon, a week in Tokyo with flights and a hotel, two weeks across Italy, a family beach holiday, a Highlands road trip, a luxury honeymoon, a business trip with a free day, one heavily edited Mexico City trip, and two conversations that never became a saved trip.

Google Gemini list prices as of October 23, 2026 (standard paid tier). We are on Google’s startup program, which covers our model bill today, but every figure here is at list price so the economics hold without it.

That is a quarter of the $0.60 in her example. The most expensive trip, two weeks across three Italian cities with 14 rounds of conversation, cost 34 cents.

Board question 1

Which parts really need a model?

Fewer than you might think. Live flight and hotel prices, availability, maps, place details and saving the trip are ordinary code and data calls. They cost no tokens, though map and place lookups carry fees of their own, covered in At scale below. The model is used to understand what you asked for and to plan the days.

The planner reads more than it writes. Each planning step re-reads about 32,000 tokens of context, two thirds of which Google serves from cache at a tenth of the price. That is why a long, heavily edited trip costs more than a short one, and why our next savings come from sending the planner less, not from writing less.

Board question 2

What happens if usage doubles or the vendor reprices?

We do not have to guess at this one. Google has announced that the model we use doubles in price on January 1, 2027, and every number in this piece is shown both ways. We also ran a cautious case of 50 cents a trip, more than three times today’s cost, to cover a further repricing or heavier use.

Return on tokens here is revenue minus model cost, divided by model cost. It stays positive in every case. A doubling halves the margin on the model line. It does not break the business.

Board question 3

What is a booking worth after the searches that went nowhere?

This is the question that decides the affiliate model, so we answer it with her booking value first, then with a floor. The programs open to us pay 4% to 5%, and the average online hotel booking is about $312, so our floor is a commission of about $12. No affiliate program is live for Just Travel yet, so both are modeled revenue.

The break-even rate already pays for every dead end, because model cost is counted on every trip, booked or not. Her example assumes one trip in ten books. At her $30 we need one in 197. Even at the $12 floor we need one in 79, or one in 43 from 2027.

Our benchmark had few dead ends, because we were running planned trips. Real visitors abandon far more. In real traffic so far, only 3 of 18 conversations became a saved trip, which puts the all-in cost at about 45 cents per saved trip today. That sample is too small to lean on, but it points the same way: even at 45 cents and the $12 floor, one saved trip in 27 has to book.

Board question 4

Does it still pay when plans change, or people book elsewhere?

Changes are already in the numbers. The Mexico City trip was edited several times and the Italy trip was rebalanced across cities. Each rewrite regenerates the itinerary at about 3.6 cents, so a heavy edit costs cents, not dollars.

Booking elsewhere is the real risk for a commission-only model, and it is why we do not rely on commission alone. A subscriber pays the same whether they book through us or not.

Who pays: three layers

Subscribers

We plan to offer Pro at $5 a month. It is not on sale yet, so treat this as planned pricing. A subscriber who plans two trips a month leaves $4.25 after the card fee and model cost. The card fee, 45 cents, is larger than the model cost of both trips. On model cost alone, they would have to plan 30 trips a month before we lost money, or 16 at 2027 prices. Counting every cost at scale, the limit is about 3 to 7 trips a month, still above the two we expect.

Bookings

Every trip, free or paid, can earn an affiliate commission when the traveler books through a link. This is the layer her piece is about, and the charts above show it works at her booking rate and at much lower ones.

Sponsored results and ads

Sponsored results are coming, and we want to be plain about how. Every sponsored result will say it is sponsored, every time. Her concern that a paid agent stops being neutral is a fair one. Our answer is to label every paid placement and leave the choice with the traveler.

Ads are a small top-up, not a business. A saved trip takes about five rounds of conversation. At one ad slot per round and a mid-range display rate, that is about one cent per free trip (a quarter of a cent to 2.5 cents across the range). Bookings carry the free tier, not ads.

At scale: every cost, not just the model

We are small today, so the fair next question is what happens at 100,000 or a million trips a month. We priced it from what we measured: model cost per trip, server time per round of conversation, and every map, place and travel-data request a trip makes. All at list prices, with no startup credits.

The higher figure in each range is the request count we measured. The lower one assumes our place cache answers 60% of map and place lookups, since popular places repeat across travelers. That share is an assumption, not a measurement.

Three things stand out. The model stays at about 16 cents a trip at any size. Servers barely register: a million trips a month runs on 12 servers across three regions for about $1,100 a month. And the largest line is map and place lookups. An itinerary looks up every stop and the route between them, and those are paid per request.

On her numbers, the answer holds at scale. At $30 a booking, one trip in 18 to 49 has to book to cover every cost, against the one in ten she assumes. Every case clears it with room to spare.

We also ran a floor. The commission rates open to us today put a hotel booking nearer $12. At that floor and the request count we measured, bookings alone fall short at 100,000 trips a month and break even at a million; with the cache assumption they clear it. That is the case subscribers are there for, and why the business does not rest on commission alone.

What is measured, and what is assumed

The model figures count model calls only; maps, travel data and hosting are priced separately in At scale. The map and place request count comes from a single session in destinations nobody had planned before, so it may move either way as we measure more.

What comes next

• Send the planner less. Re-reading context is most of the bill, so trimming it saves more than anything else.
• Smaller models for easy steps. Simple helpers already run on a lighter model. More steps can follow.
• Measure real traffic. As more travelers use Just Travel, we will publish the real-traffic figure next to the benchmark.
• Turn on revenue, in the open. Affiliate links, the Pro plan and labeled sponsored results, each reported against these numbers.

So who pays for the free travel agent? Bookings, subscribers and clearly labeled sponsors. Her questions are the right ones. The model costs about 15 cents a trip, and one booking in 197 trips covers it at her $30, or one in 79 at our $12 floor. Counting every cost, a trip at scale costs well under the $3 her own example earns from it.