Model the margin on a 20 dollar per month plan with 300 interactions at your cost per call.
Verdance Fitness charges $20 a month for MotionCue, a chat coach that writes today's workout and checks a squat or a plank from a photo or short video when a subscriber asks. An active subscriber sends about 300 messages a month, roughly ten a day. Bess Aldiss, Verdance's finance partner on the product, was asked a version of this exact question in a real pricing review: model the margin on that plan, at MotionCue's real cost per call, not a guess. The number she found at launch looked fine. The number she found a year later did not look fine on every subscriber, only on the ones who had stuck around the longest.
The direct answer
At MotionCue's real cost per call, a new subscriber's 300 calls cost about $2.18 a month against $20 of revenue, near an 85 percent margin. Left uncapped, the history sent back on every call keeps growing with a subscriber's tenure, so that same subscriber's 300 calls can cost past $4 a year in, and past $5.70 for the heaviest form-check users, pulling margin down toward 70 to 75 percent and lower still. Cap how much history rides along on each call instead of chasing a cheaper model, because the model choice was never what was moving the number.
Do this, in order
Cap the workout history sent on every call instead of letting it grow with tenure.Why: history growth is what turns an 85 percent margin into something closer to 70, and nobody chose that number on purpose.
Cost every call by type, not one blended guess.Why: a form check with a photo costs about half again as much as a plain chat, and averaging them hides which one is actually driving the number.
Quote a range to leadership, not one clean number.Why: a single figure hides how much a small assumption, like a year of history, can move it.
Track cost per subscriber by tenure cohort, not one portfolio-wide average.Why: a blended average can sit still while the longest-tenured subscribers are quietly the least profitable ones.
Keep the mid-tier model even though a cheaper one would cut cost per call by most of two dollars.Why: the cheaper model missed real form errors in testing, and correct form feedback is the actual thing the $20 is paying for.
Check the total AI cost against something real, like a single trainer session.Why: a raw dollar figure means nothing until it's compared to a price the reader already knows.
How to answer this, stage by stage
Nobody is grading whether you know a number for MotionCue. They're grading whether you know a cost per call is never really one number, it's a number attached to a moment in a subscriber's life, and averaging over that moment is how a healthy-looking margin hides a real problem.
1
Scope it to one plan, and name who owns the number
Say it like this
"Let's ground this in one real plan. MotionCue is Verdance Fitness's chat coach, twenty dollars a month, unlimited plan questions and photo or video form checks. Bess Aldiss owns the unit economics on it. First thing worth saying: three hundred interactions a month is about ten messages a day, so whatever number I give has to hold up under a real daily habit, not someone who logs in twice a week."
Why this works
Naming the owner and the real usage pattern up front stops the estimate from quietly being about a light user nobody will actually see.
2
State the equation before naming a single figure
Say it like this
"Cost per call is input tokens times the model's input price, plus output tokens times its output price. A plan question is mostly text. A form check adds a photo or a few video frames on top of that same text, which is why it costs more. Multiply the blended cost by three hundred, add the card processing fee, and subtract all of it from twenty dollars. That's the whole margin."
Why this works
Saying the equation first stops the estimate from becoming a single vague word, cheap, standing in for real arithmetic.
3
Reframe what the estimate actually needs to protect
Say it like this
"The real question isn't whether twenty dollars covers three hundred calls today. It's whether it still covers three hundred calls a year from now, once MotionCue is sending a full year of logged workouts back to the model on every single message, because that's what makes the coaching feel personal instead of generic."
Why this works
This is the moment that separates a real margin model from a single snapshot that quietly assumes every subscriber stays brand new forever.
4
Give the one decision
Say it like this
"Here's what I'd actually do: cap the history MotionCue sends on every call, instead of letting it grow with a subscriber's tenure. At a new subscriber's cost per call, margin sits near 85 percent. Left uncapped, that same subscriber's calls can cost twice as much a year in, and the cap is what keeps the number honest, not the model choice."
Why this works
This is the concrete, defensible decision, not a wish that margin could just stay wherever it started.
5
Own the numbers, and show a range, not one figure
Say it like this
"A plan-only message costs about two thirds of a cent to answer, $0.0066. A form check with a photo costs about a cent, $0.0099. Blend those at a new subscriber's mix and three hundred calls run about two dollars eighteen a month. Blend them at a year-in subscriber's mix, more history, more form checks, and the same three hundred calls run past four dollars twenty. The honest range is about two dollars to just under six, not one number."
Why this works
Giving both ends of the range, and saying what pushes each end, tells the interviewer this wasn't one lucky guess.
6
Sanity check it against something real, and name what you turned down
Say it like this
"Here's the check that mattered: even the worst case, about five dollars seventy in model cost, is less than what Verdance charges for one fifteen-minute call with a human trainer, about twenty eight dollars. We looked at a cheaper model that would have cut that cost by most of two dollars. We turned it down, because it missed real form errors in testing, and correct form feedback is the actual thing subscribers are paying twenty dollars for."
Why this works
Comparing the dollar figure to something the interviewer already understands, and naming the rejected option, makes this a real decision instead of a number floating with nothing to hold it down.
7
Say which assumption swings it most, then close on one line
Say it like this
"Two assumptions swing this the most, and they're not the same kind of lever. History growth swings about two dollars a month, the biggest single move in the whole model, and nobody chose that number on purpose. The model tier is close behind, about a dollar seventy five, but that lever is already spent, on purpose, to protect quality. So: cap the history, cost each call by type, and watch margin by tenure cohort, not by one blended average that can hide your best customers quietly getting the most expensive."
Why this works
Closing on the actual decision, in one breath, is what makes the answer sound rehearsed instead of like a story that trailed off.
Let's learn
MotionCue is the chat inside Verdance Fitness's app. Ask it what to do for today's workout, or send it a short video of a squat, and it writes back a plan, or tells you what's off in your form, in a few seconds.
Five steps. The third one, the model call, is the one the whole margin question actually gets built around.
Before MotionCue, the same twenty dollars bought a library of workout PDFs that never changed once you downloaded them. A real person checking your squat cost about twenty eight dollars for a single fifteen-minute video call on Verdance's trainer marketplace, and most subscribers never booked one.
Twenty dollars, a static plan nobody read twice, and a trainer call almost nobody actually booked.
Now the same twenty dollars buys a coach that answers back, instantly, about ten times a day for a subscriber who actually uses it. Three hundred messages a month.
Knowledge spark: what does a chat coach actually send the model?
Every message carries more than what you typed. It carries your profile, your recent workouts, and, for a form check, a photo or a few frames from your video. All of that gets read by the model every single time, and all of it costs money to read.
The turn: the extra messages are not the risk. Ten a day, answered instantly, is exactly the habit Verdance wanted to build. The risk is what those messages quietly cost to answer, because MotionCue sends a subscriber's whole logged history back to the model every time, so the coaching feels personal instead of generic. A brand new subscriber's messages cost a fraction of a cent each. A subscriber with a year of logged workouts sends a much bigger message every time, and nobody had priced the plan around what a loyal subscriber actually costs.
We did not price the twenty dollar plan against a subscriber. We priced it against a subscriber's first month.
Blended cost per call, by months since signup
Near launch numbersDrifting, nobody alertedA year in, uncapped
No single jump. A steady climb, one tenth of a cent at a time, invisible inside a blended portfolio average that kept looking about the same every month.
At its worst, the least profitable subscriber on the twenty dollar plan is not the one who barely uses it. It's the most loyal one: a year in, sending ten messages a day, more of them form checks, with a full year of history riding along on every one. That subscriber's true monthly cost can climb past five dollars seventy, on the same three hundred calls that cost a new subscriber about two dollars eighteen.
The decision that mattered
MotionCue's cost model was built and signed off against a new subscriber's usage, in the two weeks after signup. Nobody set a separate check for what the same subscriber would cost to serve a year later.
What I would leave alone: unlimited messaging itself. Capping how many messages a subscriber can send would have fixed the margin and broken the actual product. Ten messages a day was never the expensive part. The growing history riding along with each one was.
The lesson: cost per call is never really one number. It's a number attached to a moment in a subscriber's life, and if you don't watch how that moment changes over a year, the average will lie to you right up until your best customers are the ones quietly costing you the most.
Now here is the same thing as a story
Read the long version below when you want to feel why a margin that tested perfectly in week one could still be quietly wrong by month twelve, one loyal subscriber at a time.
Bess Aldiss had spent two years at Verdance building the numbers behind every plan the company shipped, mostly plain arithmetic: printing costs, payment fees, the marketplace cut on a trainer's session. When MotionCue needed a cost model before launch, Bess did what she always did. She built the number from a real week of usage and trusted what she saw.
What she tested against was the first two weeks of MotionCue's private beta: forty employees using it exactly like a brand new subscriber would, mostly plan questions, the occasional photo. Against that group, the blended cost came out to about two dollars eighteen a subscriber a month. Bess signed off on the twenty dollar price with a margin near eighty five percent, and told the launch team the number would hold.
The first three months after launch looked exactly like the beta. New subscribers signed up, sent mostly plan questions, and the finance dashboard's blended cost stayed steady, right where Bess had modeled it. The margin line on her monthly deck sat near eighty four percent for three straight reviews, and she moved on to pricing the next feature.
Nothing broke. No single subscriber's bill spiked, no alert fired, no support ticket said the app cost too much. It built up two or three tenths of a cent at a time, spread across every subscriber's whole tenure, invisible inside one blended, portfolio-wide average that kept looking almost exactly the same month after month. It took a routine quarterly review, the kind where someone finally breaks a number out by signup cohort instead of blending it, for Bess to notice the line for twelve-month subscribers didn't match the line for one-month subscribers at all.
No single bad month. A slow climb nobody's dashboard was built to catch, until a routine cohort review finally split the average apart.
She pulled the actual call logs for a hundred year-long subscribers and matched them against a hundred new ones. New subscribers: about two dollars eighteen a month, same as the beta. Year-long subscribers: past four dollars twenty, some well past five seventy, on the exact same three hundred calls. It took her most of a day, reading through raw request payloads line by line, before she found why: MotionCue was sending a subscriber's entire logged workout history back to the model on every single message, and a year of logged workouts is a lot more to read than a first week's worth.
The wrong number on the dashboard was never the real cost. The real cost was twelve months of pricing decisions, hiring plans, and a board deck, all built on a margin nobody had checked past week three.
We did not lose margin on a bad subscriber. We lost it on our best one, one loyal month at a time.
It was never really about the exact dollar figure. Bess never had one true number to defend. She had a model that was quietly a different number for every subscriber depending on how long they had stayed, and averaging them together hid that completely.
The decision that opened the door went back to a fifteen-minute engineering sync, three weeks before launch. Sending the full history was the obvious way to make early coaching feel personal instead of generic, and it worked, beta testers said MotionCue's plans felt like they actually knew them. Nobody in that meeting decided the same history should keep growing, forever, with no ceiling, for as long as a subscriber stayed.
The old design wasn't careless. It was signed off honestly against a subscriber's first two weeks, and simply never re-checked past that.
Run the same year again, with a rolling summary capping the history at a fixed size from day one instead of the full log. A year-long subscriber's calls still cost more than a new subscriber's, that part was never going away, but the number holds near three dollars instead of climbing past five seventy, and the quarterly cohort review finds a gap worth a footnote instead of a gap worth a re-forecast.
One design let the first three weeks of beta data decide what every subscriber, for their whole time on the plan, would quietly cost. The other let the cap decide it, on day one, and left the beta data where it belonged: a starting guess, not a permanent ceiling nobody remembered to check.
What Bess would tell herself, back in that first meeting: sending the full history wasn't wrong. It just needed a second promise sitting right next to it, a cap on how big that history could get, and nobody had written that promise down.
BOUND, or how Bess turned three hundred chats into a number she could defend
Not a story dressed as a framework. This is an estimation problem with a real gap hiding in it, and BOUND is what turns "the model is cheap" into an actual, defensible margin.
BBreak it down. What's the actual equation?
MotionCue's monthly cost is two things multiplied, added up, then set against the price. Cost per call equals input tokens times the model's input price, plus output tokens times its output price, blended across plan chats and form checks. Multiply that blended cost by three hundred calls, add the card processing fee, and subtract the total from twenty dollars. Two terms per call, one multiplier, one price. Not one number standing in for "the AI costs something."
Say the equation before naming a figure, or the margin quietly becomes whatever the finance dashboard's headline number happened to say that quarter.
Two of these terms barely move. One of them, the history riding along in input tokens, is the one that quietly changes with every month a subscriber stays.
OOwn the numbers. Where did each one come from?
A plan-only message: about 950 input tokens, 420 output tokens, at the mid-tier model's rate of $2.50 per million input tokens and $10 per million output tokens, pulled from the vendor's own rate card. That prices out to about $0.0066. A form check adds a photo or a few video frames, about 760 tokens worth, and runs a little longer on the output, 560 tokens: about $0.0099. This is also where the rejected alternative sits: a cheaper mini-tier model, priced near a fifth of that, dropped after it missed real form errors in testing that the mid-tier model caught.
Owning the number means saying where it came from and what it would have cost to buy it cheaper, not just stating a figure.
UUse a range, not one number.
A new subscriber's mix, about 80 percent plan questions, 20 percent form checks, blends to about $0.0073 a call, near $2.18 for three hundred calls. A year-long subscriber, more history riding along on every message, a mix that shifts toward more form checks too, blends closer to $0.0142 a call, past $4.26 for the same three hundred calls. Push it further, a veteran subscriber doing mostly form checks, and it can run near $0.019 a call, close to $5.70. The honest range is about two dollars to just under six, not one clean number.
The whole argument for tracking margin by tenure cohort, not one blended figure, lives inside that range.
The build-up: a new subscriber's month, against a year-in subscriber's month
AI cost per callCard processing feeMargin left
Same $20, same 300 calls. The margin bar shrinks from $16.94 to $14.86 on the exact subscriber Verdance least wants to lose, purely from history riding along on every call.
The whole cost range sits under two cents a call. The trainer session, marked for scale, sits nowhere near it.
NNail the sanity check. Does the number survive being compared to something real?
Even the worst case, $5.70 in model cost a month, is less than what Verdance charges for one fifteen-minute video call with a human trainer, about $28. Every case here, from $2.18 up to $5.70, stays under 30 percent of the $20 price, so gross margin never drops below about two thirds even in the worst case, still healthy for a consumer subscription. The number that should worry Bess isn't any single figure on this page. It's that the blended, portfolio-wide average sat near $2.20 for a year while individual year-long subscribers were already past $4.
The hardest step, and the one most answers skip. A number that sounds reasonable in isolation can still be sitting on top of a real, quietly widening gap.
DDirection. Which assumption would move the answer most?
Two assumptions swing this the most, and they're not the same kind of lever. History growth, new subscriber to year-in, is the single biggest swing in the whole model, about $2.08 a month, and it's the one nobody chose on purpose, it just happens for as long as a subscriber stays. The model tier is close behind, about $1.75, but that lever is already spent, deliberately, to protect quality. So the direction that actually matters isn't which one moves the number most on paper. It's which one nobody is watching.
Naming the assumption nobody is tracking, not just the one with the biggest number attached, is what a good estimator does that a bad one skips.
What moves MotionCue's monthly cost, and who chose it on purpose
The model swap looks like the obvious lever. It's already spoken for, on purpose, for quality. The bar that actually needs a fix is the one at the top, because nobody assigned it an owner.
Three things worth stating directly, since this is where the real judgment sits. The AI-specific failure worth naming is a kind of quiet cost drift: MotionCue's blended average looked stable because new subscribers kept joining and diluting it, even as every individual subscriber's true cost climbed the longer they stayed, the same shape as a metric that looks healthy on a dashboard while a real segment underneath it quietly gets worse. The guardrail is tracking cost per active subscriber by signup cohort, not one blended, portfolio-wide number, so a rising line for twelve-month subscribers shows up long before it drags the whole average anywhere. The threshold isn't a hard, always-true rule either: MotionCue's margin model holds as long as the ninetieth percentile tenure cohort stays under about two cents a call, most months, not that every single call must clear some fixed cap. And the trade being accepted plainly: Verdance is paying more per call to keep the mid-tier, vision-capable model instead of a cheaper one, in exchange for form feedback subscribers can actually trust, and it is capping history instead of capping messages, trading a slightly less personalized coach for a margin that survives a subscriber's second year.
And if you want to be sure it really works, try it somewhere else
Same five letters worth repeating, a symptom-triage chat instead of a workout coach, and this time the cost swing doesn't creep in slowly. It arrives all at once, on a bad week.
TailLine is the chat inside Whitmere Veterinary Group's app: describe a symptom, or send a photo of a limp or a rash, and a vet-trained assistant triages whether it can wait for a morning appointment or needs an emergency clinic tonight. Fifteen dollars a month, about 150 chats included. Dorian Ilic, the engineer who owns TailLine's inference budget, ran the same margin model Bess did.
The decision Dorian would take back
TailLine's first version priced every chat the same, whether it was a plain question about flea medicine or a photo of a swollen paw. One flat number, no split by type.
A vet chat has almost no tolerance for averaging by mistake. A plain question runs about the same cost as MotionCue's plan chat, well under a cent. A photo triage runs close to three cents, because a real triage call has to reason carefully about what it's looking at, and get it wrong far less often than a human vet on call would tolerate. The risk for TailLine was never a slow, gradual drift like MotionCue's. It arrives in spikes: a hot week with a lot of limping dogs and swollen paws can push a fifth of that week's chats to the expensive kind, all at once, on subscribers who did nothing different, they just happened to need Whitmere the week something was actually wrong. A normal week blends to about $0.008 a chat, near $1.20 for the month, margin over 90 percent. A spike week, where the photo-triage share jumps past a fifth of all chats, can push the blend past $0.016 a chat, near $2.40, margin still healthy, near 84 percent, but the swing happens inside days, not months, so a single monthly average can hide an entire bad week.
MotionCue's swing hides inside a year. TailLine's swing hides inside a week, which means a monthly average is already too slow to catch it.
Same method, different lever: for MotionCue, the swing built slowly, inside one subscriber's own tenure, and the fix was a cap that grows with time. For TailLine, the swing arrives all at once, across a whole cohort in the same bad week, and the fix isn't a cap on any one call, it's costing plan chats and photo triages separately so a bad week's mix shows up in the number immediately instead of waiting for a monthly average to catch up.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: break the cost into its real types first, price each one, then multiply by volume and subtract from the price, never one blended guess.
Cost: if Whitmere's inference budget got cut, the flat fifteen dollar price wouldn't move; the free-tier limit on photo triages would tighten instead, protecting the number that actually swings, not the one that barely does.
The model got better, for real: say the next model reads a photo triage twice as accurately for the same price. That doesn't touch TailLine's price at all, it just means the same margin now comes with a triage a subscriber can trust more, which is the whole point of paying for the mid-tier model in the first place.
Where people run it wrong.
They price the plan off a single blended cost per call, calculated once at launch, instead of once per call type.
They watch a monthly average and miss a swing that arrives and resolves inside a single week.
They cut the model tier to protect margin, without checking whether the cheaper model was actually good enough at the thing subscribers are paying for.
How to use it live. Ask the type-split question before quoting any number: is this cost the same for every call, or does it split by what the call actually does, a plan question versus a photo triage, a new subscriber versus a year-long one? On paper those look like the same three hundred calls. In the real cost, they almost never are.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework is this, and what's its one job?
Tap to flip
ANSWER
BOUND: show the arithmetic, own the assumptions. Built for estimation and margin questions like this one, not a story about a habit.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Bess Aldiss, the finance partner who owns MotionCue's unit economics at Verdance Fitness, and had priced every plan the company shipped before MotionCue existed.
3 · THE BLIND SPOT
What did the 85 percent margin get tested against that real subscriber tenure didn't match?
Tap to flip
ANSWER
A two-week beta of forty employees acting like brand new subscribers. It held there. It fell toward 74 percent, and lower for the heaviest users, once real subscribers stuck around a year.
4 · THE EQUATION
What two terms make up MotionCue's cost per call?
Tap to flip
ANSWER
Input tokens, profile, history, and any photo, times the model's input price, plus output tokens times its output price, blended across plan chats and form checks.
5 · THE OLD DECISION
What decision would Bess take back?
Tap to flip
ANSWER
Sending a subscriber's whole logged history back to the model on every call, with no cap, instead of a rolling summary, because it made early coaching feel personal and nobody priced what it would cost by month twelve.
6 · THE NUMBER
Fill in the blank: a new subscriber's 300 calls cost about $___ a month; a year-in subscriber's same 300 calls can cost about $___.
Tap to flip
ANSWER
About $2.18, and about $4.26, climbing toward $5.70 for the heaviest, longest-tenured form-check users. Uncapped, the number keeps climbing the longer a subscriber stays, exactly backwards from what a healthy margin should do.
7 · THE REPLAY
Same year, new design, what changes?
Tap to flip
ANSWER
With history capped from day one, a year-in subscriber's cost holds near three dollars instead of climbing past $4.26, and a routine cohort review finds a small footnote instead of a number worth re-forecasting the whole plan around.
8 · CROSS-PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what's the different lever there?
Tap to flip
ANSWER
TailLine, a symptom-triage chat for Whitmere Veterinary Group. The lever there isn't slow drift over a subscriber's tenure, it's a sudden weekly spike in photo triages that a monthly average can hide.
Check yourself Score: 0 / 0
True or false
1. True or false: MotionCue's blended, portfolio-wide average cost per call stayed roughly flat for a year because the cost of serving any one subscriber stayed flat too.
True
False
Show hint
Check the trigger paragraph in the story, and the N step in the framework recap.
Show answer
False. The portfolio average stayed flat because new subscribers kept joining and diluting it. Any one subscriber's own cost kept climbing the longer they stayed; the average just never showed it.
Fill in the blank
2. MotionCue prices its model at $___ per million input tokens and $___ per million output tokens.
Show hint
Look at the O step in the framework recap.
Show answer
$2.50 per million input tokens, and $10 per million output tokens. Pulled from the vendor's rate card for the mid-tier, vision-capable model Verdance runs, the tier chosen over a cheaper mini-tier for quality reasons.
Multiple choice
3. Per the sensitivity numbers in the framework recap, which assumption swings MotionCue's monthly cost the most?
A. Whether the payment processor's fee changes
B. History growth, from a new subscriber to a year-in subscriber
C. Whether the plan is billed monthly or yearly
D. The exact wording of the system prompt
Show hint
Check the U and D steps, and the horizontal bar chart right after them.
Show answer
B. History growth swings about $2.08 a month, the single biggest lever in the model, and the only one nobody assigned an owner to watch.
Short answer, name the rejected alternative
4. What alternative did Bess's team consider and reject, instead of capping the history sent per call?
Show hint
Look at the O step, and stage 6 of the walkthrough.
Show answer
Model answer: Swapping to a cheaper mini-tier model. Rejected because it missed real form errors in testing that the mid-tier model caught, and correct form feedback is what the twenty dollar price is actually paying for.
Short answer, apply it yourself
5. Think of a subscription you pay a flat monthly rate for. Does it likely cost the company roughly the same to serve you in month one as it does after a year of using it? Name one reason it might not.
Show hint
Think about whether the service keeps more data, more history, or more state about you the longer you stay.
Show answer
Model answer: A cloud photo-backup app likely costs more to serve a long-time subscriber, since it's storing and syncing years of photos instead of a few weeks' worth, even though the monthly price never changed.
Multiple choice
6. Per the sanity check in the framework recap, how does even the worst-case model cost compare to something Verdance already charges for?
A. It's about triple the cost of a single trainer session.
B. It's less than the cost of one fifteen-minute human trainer call.
C. It's roughly equal to the full $20 subscription price.
D. There's nothing to compare it to, since AI costs are unprecedented.
Show hint
Check the N step in the framework recap.
Show answer
B. The worst case, about $5.70, is still less than the roughly $28 Verdance charges for a single fifteen-minute trainer call, the comparison that makes the raw dollar figure mean something.
Before you close the answer
Why this works
Tests whether you'll track a real cost by who's actually generating it, cohort by cohort, or trust one blended average because it happens to look calm on a dashboard. Most candidates stop at the launch-week number.
Follow-up traps
"Why not just cap total messages instead of capping history?" Response: ten messages a day was never the expensive part, the growing history riding along with each one was. Capping messages would have broken the actual product to fix a cost that message count never caused.
"Couldn't you just raise the price to twenty five dollars and stop worrying about it?" Response: that fixes today's margin and hides the same drift for another year, since the uncapped history keeps growing either way. The cap is what keeps the model's own behavior from being the thing quietly setting the price.
If pressed
The rolling summary that replaced the full history compresses everything older than the last two weeks of workouts into a short structured recap instead of the raw log, which is what keeps a year-in subscriber's input tokens close to a new subscriber's, not just smaller than uncapped.
From U2xAI Academy
From answering questions to owning outcomes.
A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.
