Benvenuti, amici miei. (Or, welcome my friends in Italian). I’m taking a trip to Italy this summer, and like some of you, I like to learn a few phrases so that I don’t seem like a lame tourist. And more importantly, to make sure that I order the right pasta.
I’ve tried practicing with various products, including Duolingo, which I’m sure everyone has tried at one point or another. It’s fun, it feels like a mobile game, and they get bonus points for their silly mascot. But language learning requires actual conversations, not just vocab lessons. That’s why I’ve found better success using an online tutor for long-term language learning, but that seems like overkill for a vacation.
I think Duolingo recognizes that as well. They now offer a video call feature where you can chat with one of their characters using AI, and I’m intrigued. This week, I’m diving deep into this feature and the strategy that might change the face of language learning.
P.S. If anyone at Duolingo is reading this, please add Punjabi as a language option. I could use the practice.
TL;DR
The thirty-second version of what Duolingo did.
Long ago, in a galaxy far, far away, Tesla sold exactly one car: the Roadster. It cost about $98,000, they only made 2,450, and it was never meant to be the business. It was a proof of concept, a glimpse of what an electric car could be, sold to a handful of early adopters with money to burn. The point was to fund the technology at the top, then ride the cost down to the masses. Duolingo took a similar approach with their feature, Video Call with Lily.
This looks more like a Ferrari EV than the actual new Ferrari EV
In March 2023, Duolingo launched AI features behind a new $30-a-month paid plan called Max, including:
Explain My Answer - an in-lesson feedback feature
Video Call with Lily - a conversation feature
By February 2026, Explain My Answer was free for everyone, and Video Call had started sliding down to the $13 Super tier, one rung below Max.
Duolingo gated these features because they were brutally expensive to serve in 2023, not because they wanted to wall them off forever. The premium tier funded the experiment and proved the features were winners: over the past year, the average words spoken per user in Video Call more than doubled, and Duolingo's own efficacy studies showed Video Call measurably improved speaking and confidence, not just engagement.
Per-unit AI costs kept falling, so Duolingo could offer the feature to far more learners while its Q1 2026 gross margin actually rose to 73%, up nearly two points year over year. The $30 feature from 2023 was cheap enough by 2026 to move down a tier, allowing them to increase the feature reach from the roughly 1.1 million people who paid $30 for Max to the roughly 11 million on the $13 Super tier, about ten times the audience.
This works because Duolingo has two paid tiers: Max and Super. Most apps only have one paid tier. Those apps can’t move their AI features to a lower paid plan to increase reach when costs go down.
The Shift: Premium → Default.
Inference costs fall on a predictable curve. The strategy is to gate expensive features at premium today, then migrate them down the tier ladder as costs drop, while seeding the next too-expensive bet at the top.
SPONSORED BY
You bought something online with a credit card. You may have even used the “right” credit card.
But did you:
Activate offers on your credit cards
Use a shopping portal
Check for heavily discounted gift cards
Shoppers who use Savewise earn 7-10x more points & miles for things they’re already buying.
Savewise finds every way for you to earn rewards and puts them all in one place.
Under The Hood
The three strategies that make it work.
1. Test expensive new features on premium customers
Pricey business-class seats subsidize the economy cabin so the rest of us peasants can still get where we're going. Hardcover books come out first for the die-hards who'll pay double, then the paperback lands a year later. I foot the Netflix bill so my whole family watches for free. The market is littered with a premium segment quietly underwriting a cheaper one.
Duolingo runs the same play, with one major difference: the goal is to make a forward bet, not to permanently gate a feature.
Max launched in 2023 with the release of GPT-4 to ride the new wave of LLMs. The flagship feature was the Video Call with Lily. It was the first real shot at simulating a real conversation on the platform. Lily kicks off the conversation, but you can talk about anything that you want. It helps you practice speaking in a low-pressure situation that can otherwise be nerve-wracking.
Apparently, Lily’s sassy personality shines through with “her signature sarcasm”. Based on the demo, it feels more like you just woke her up from a nap and she can’t be bothered to chat. So, pretty much how any local feels hearing a tourist butcher their language while asking for a cup of coffee.
But it works really well. In studies, learners who used the feature regularly showed greater improvements in speaking than those who stuck to regular lessons. On the money side, Max grew to roughly 9% of paid subscribers by late 2025, up from 7% a year earlier, a real signal that people will pay 3x Super's price for AI conversation practice.
2. Distribute through the tiers once AI costs drop
Once a feature is proven and the cost of serving it drops, Duolingo pushes it down through the tiers to expand reach. But where it lands depends on how expensive it still is to run as costs come down.
Take Explain My Answer, the feature that gives you personalized explanations of the grammar and vocabulary behind a mistake. It launched Max-exclusive in 2023, and it's now free for everyone. It's cheap enough to serve at scale that paying subscribers comfortably cover the free tier's usage.
Video Call with Lily is a heavier lift. Simulating a live conversation costs far more to run than spitting out a grammar explanation, so it isn't a free-tier feature. It's taking one step down the ladder, from the $30 Max tier to the $13 Super tier. The rule of thumb: the cheaper a feature is to serve, the further down it travels, and the cost curve decides when. Maybe one day the video calling feature can be free for all if costs drop low enough.
This strategy only works if you have two paid tiers. If you have one, you can't move AI features down the ladder while keeping them paid. Your only option to widen reach is making them free, which kills revenue and gives subscribers a reason to cancel.
3. Find another bet ahead of the cost curve, and repeat
Distributing the winner clears the way for the next expensive experiment. Max isn't a fixed set of features. It is a permanent slot for whatever's too expensive to give away right now. So the cycle never stops: today's premium bet becomes tomorrow's default, and a new bet takes its place at the top. Duolingo said as much in its Q4 letter, promising to keep experimenting with what lives in Max.
On January 14, 2026, Duolingo dropped a new feature called Video Call with Falstaff into Max. Six weeks later, on February 26, it told shareholders that Video Call with Lily was moving down to Super. A new expensive bet went in at almost the exact moment the proven one started leaving.
The key is to keep pushing the limits of what is technically and economically feasible. Video Call with Falstaff goes further than Lily: it coaches you in real time, gives feedback in your own language, and hands you phrases and translations when you get stuck. It's a more complex, more ambitious feature than the original and can only scale if model costs continue to drop in the future.
The Model Y is now Tesla's best-selling car. Video Call with Lily is now $13 a month. Both exist because someone paid the premium price first.
What I’d Ship
The Duolingo strategy works well for any product that has two or more paid tiers and can designate one of the tiers for testing AI features. For products that only have one plan, like the two mentioned below, this strategy still works as long as you create a new ultra-premium offering.
Strava
I’m a newly minted runner. And honestly, what’s the point of running if you’re not logging it on Strava to share with your friends? Strava sells one thing: a subscription, around $12 a month, sitting on top of a free tier. And it already has an AI feature, Athlete Intelligence, which reads your run and hands back a summary of how it went.
I'd ship a tier above the current subscription with a real AI coach that builds your training plan, adjusts it when you skip a day or show up wrecked, and talks you through the hard sessions in real time. It feels like their Runna acquisition was made for this.
That's expensive to run (ba-dum-tss) today, which is exactly why it belongs at the top, where the most committed athletes will fund it. As the cost drops, the coach slides down into the regular subscription, and the basic summaries drift toward free as the hook.
Headspace
This app is a bit different. Headspace isn't freemium like Duolingo or Strava. There's no real free tier, just a trial and then one subscription, about $70 a year, with the AI companion Ebb baked in. So its ladder is shorter and the bottom rung isn't free, it's that paid plan.
If Headspace wants to give Ebb a face: real-time video, a companion you can actually look at while you talk through a panic attack. That's an order of magnitude more expensive to run than voice, and there's no safe way to drop it into the $70 plan and hope the math holds.
With one tier, that feature just doesn't get built. They need a tier above the base plan whose whole job is to trial the expensive next thing.
SPONSORED BY
Built by yours truly → a 2-minute pre-game routine for high-stakes moments:
The important job interview in 5 minutes.
That big board presentation starting in 10 minutes.
The hard conversation you’ve been putting off with your team.
"Just breathe" doesn't work when you're already spiraling, and Calm and Headspace aren't built for the panic before the big moment.
Primo is. It walks you through a 2-minute routine using science-backed techniques from Stanford and Harvard.
Extra Credit
For my fellow nerds out there who want to know how they’ve designed the system prompts for the LLMs powering the calls. I don’t want to spoil the article, but I learned that there is a separate LLM call and instruction set for the first question. Apparently, that opening question is formulated while the Video Call is ringing.
It’s as secret as the In-N-Out secret menu. A manifesto written by Elon before the release of the first Tesla car. “Almost any new technology initially has high unit cost before it can be optimized” → this is exactly part of the core thesis of the Duolingo strategy. It’s also a great read for anyone who likes cars.
This one was super fun for me since Duolingo has always been the gold standard for consumer products.
If you found anything in this email interesting or learned anything new, please do me a huge favor and forward it to one person who'd find it useful!
— Amaraj (aka the man who muted his Duolingo notifications)
The Meme
Everyone should fear Duo