From Shubhi K | Product & Market Analysis
Duolingo Gross Margin at 72.9%: How Much Came From AI Calls Under a Cent
On this page
Duolingo gross margin was 72.9% in the third quarter of 2026, up from 72.5% a year earlier. The 10-Q credits lower feature based AI costs. On the call, management said a Video Call now costs well under a cent to serve, against about $0.30 at launch. Read the filing closely and the model savings explain less of the margin than the headline suggests.
Key takeaways
- Duolingo's Q3 2026 gross margin was 72.9%, against a guide of 71.0%. The 10-Q attributes the year-over-year gain to higher subscription gross margin from lower feature based AI costs, and the beat is worth roughly $5.9 million of gross profit.
- Video Call inference fell from about $0.30 to well under $0.01 per call. The CEO credited the move to open-weight models, which is a cut of more than 96% on a per-call basis and is a management statement, not an audited figure.
- Cheaper models funded wider access rather than fatter margins. Duolingo spent the savings putting Video Call into the Super tier for roughly 10 times as many learners, and sunset its Max tier, so reported margin barely moved.
- The margin gain did not reach the bottom line. Adjusted EBITDA margin fell from 29.5% to 25.5% in the same quarter, as research and development spend rose about 30% and an $11.4 million bonus accrual landed.
Short answer
Duolingo's gross margin rose 40 basis points to 72.9% in Q3 2026. The company credits cheaper AI features, and Video Call inference fell below a cent per call after a switch to open-weight models. Most of that saving was spent widening AI access, so Q4 margin is guided down to 72.0%.
What the 10-Q says about Duolingo gross margin
The filing language is short and specific. Duolingo's 10-Q states that total gross margin increased to 72.9% in the three months ended 30 September 2026, from 72.5% a year earlier. For the nine months it rose to 72.8%, from 72.0%. It says the increase in both periods was primarily attributable to higher subscription gross margin, reflecting lower feature based AI costs.
That attribution matters because of what the company counts as cost of revenues. The 10-Q says cost of revenues predominantly consists of third-party payment processing fees charged by distribution channels. Hosting fees and third-party AI costs sit on top. Customer support, amortisation of revenue-generating capitalised software and some depreciation sit in the same line, to a much lesser extent.
So the biggest cost under the gross margin is the toll the app stores and payment processors take on each subscription. AI inference sits beside hosting as a second-tier item. Duolingo does not break the AI portion out as a number in the filing, and nothing in this post should be read as if it did.
The Q3 shareholder letter describes 72.9% as a modest improvement over the prior year and above expectations. It credits the outperformance primarily to lower feature-based AI costs and favourable hosting cost trends. Hosting appears in the letter and not in the 10-Q's one-line attribution. So the AI saving was not the only cost tailwind in the quarter.
Three months earlier, the Q2 letter guided Q3 gross margin to approximately 71.0%. The actual came in about 1.9 points higher. On Q3 revenue of $309.0 million, that gap is worth roughly $5.9 million of gross profit in one quarter. CFO Gillian Munson put it plainly on the call. The company thought margin would be about 2 points lower, and AI costs had not been as high as expected.
The Q3 unit economics walk, computed from the filing
Duolingo reports cost of revenues in dollars and user metrics in millions. Dividing one by the other gives a cost-to-serve figure the company does not publish. The table below does that arithmetic. Every input is from the 10-Q or the shareholder letter. Every ratio is my own calculation.
| Line | Q3 2025 | Q3 2026 | Change |
|---|---|---|---|
| Revenue | $271.7M | $309.0M | +14% |
| Cost of revenues | $74.8M | $83.9M | +12% |
| Gross margin | 72.5% | 72.9% | +40 bps |
| Daily active users (quarter average) | 50.5M | 64.9M | +29% |
| Paid subscribers (period end) | 11.5M | 13.1M | +14% |
| Cost of revenues per DAU, per quarter (computed) | $1.48 | $1.29 | about -13% |
| Cost of revenues per paid subscriber, per quarter (computed) | $6.50 | $6.40 | about -2% |
| Incremental gross margin on the $37.3M revenue increase (computed) | $9.1M of added cost against $37.3M of added revenue | about 75.7% | |
Per-subscriber figures divide quarterly cost by period-end subscribers, not the quarter average. That slightly understates cost per subscriber in a growing base. Cost of revenues includes payment fees, hosting, AI and support, so neither per-unit figure is an AI cost.
Cost per DAU fell, cost per subscriber barely moved
The two per-unit lines tell different stories, and the gap between them is the useful part. Cost per daily active user fell about 13%, because DAUs grew 29% while total cost grew 12%. Cost per paid subscriber fell only about 2%.
That pattern fits the cost structure in the filing. Payment processing fees scale with paying subscribers, not with free users, so they hold cost per subscriber steady. Free users add hosting and some AI cost, but far less than they add to the DAU count. My read is that cost per DAU flatters the AI story. Cost per subscriber is the more honest test of whether serving a paying learner got cheaper.
Each extra dollar of revenue in Q3 cost about 24 cents to deliver, giving an incremental gross margin near 75.7%. That is about 3 points above the reported average. When marginal revenue carries a higher margin than average revenue, the business is getting cheaper to serve at the edge. That is the outcome lower AI costs should produce.
Video Call inference cost: $0.30 to under a cent
Video Call is the AI conversation feature Duolingo announced at its Duocon 2024 event. Learners speak with an animated character called Lily. It launched inside Duolingo Max, the company's most expensive tier, because it was costly to run.
On the Q3 call, CEO Luis von Ahn said Video Call was costing about $0.30 per call in AI inference at launch. He said the company had reduced that to well under $0.01 per call, mainly by moving to open-source or open-weight models. Taking $0.01 as a ceiling, that is a reduction of more than 96% per call.
He also explained why Max existed in the first place. The company had to start it because some AI features were very expensive to serve. That is a candid admission that the premium tier was partly a cost-recovery mechanism. The economics of open-weight models are now dismantling it. The broader production trade-offs of that switch are covered in the comparison of open-weight models in production.
At $0.30, one weekly call for every paying subscriber would have cost about 61% of the quarter's cost of revenues. No subscription business at a 72% margin absorbs that without repricing. At under a cent, the same usage costs about 2% of the line. That is why the feature could move down a tier.
I would treat this as the real lesson of the quarter. The model price did not mainly raise the margin. It changed which customers a feature could be offered to. The pattern repeats across how inference costs flow into AI product margins.
Separating model savings from feature expansion
Three forces moved Duolingo's AI cost per user in 2026, and they pushed in opposite directions. The filing reports only their net effect. Separating them requires reading the shareholder letters in sequence.
| Driver | Direction on margin | Evidence | Quantified by the company? |
|---|---|---|---|
| Cheaper inference per call (open-weight models) | Up | Video Call from about $0.30 to under $0.01 per call, Q3 call | Per call only, no total |
| Video Call extended from Max into Super | Down | Roughly 10 times as many learners get access, Q4 2025 letter and Q3 call | No cost figure |
| Favourable hosting costs | Up | Named in the Q3 letter as part of the outperformance | No |
| Revenue mix shifting toward subscriptions | Up | Subscription revenue grew 18% vs 14% for the total, 10-Q | Revenue lines only |
| Max sunset and migration to Super Family | Small, guided into Q4 | Max is single digits of subscribers and bookings, CFO on the call | Included in the 72.0% guide |
The February plan assumed expansion would win
In its Q4 2025 letter, Duolingo guided gross margin to approximately 71% in Q1 2026 and roughly 69% for the rest of the year. It said the decline would be driven primarily by expanding access to AI-powered features for all users. In other words, management planned for feature expansion to cost about 3 points of margin.
That did not happen. Q2 came in at 72.6% against a 71.0% guide. The letter cited a measured pace of feature expansion and AI cost efficiencies. Q3 came in at 72.9% against another 71.0% guide. The Q2 letter had already raised the full-year expectation to about 71.6%, citing AI cost trends.
Start with nine-month gross profit of $655.0 million. Add a Q4 guide of 72.0% on $322 million of revenue. That gives about $886.8 million of gross profit on about $1,221 million of revenue, and implies a full-year 2026 gross margin near 72.6%. Full-year 2025 was 72.2%. On the company's own numbers, a year planned as a margin step-down is on track to finish slightly above the year before.
This is the cleanest way to separate the two forces. The expansion cost was real enough that management planned for it. The model savings were large enough to cancel it. Neither shows up as a line you can read directly. Buyers face the same problem when they try to find the real gross margin behind an AI feature.
Subscription mix did some of the work on gross margin
The 10-Q attributes the margin gain to subscription gross margin. But the revenue mix moved too, and that matters for anyone copying the conclusion. Subscription revenue grew 18% to $270.0 million, while total revenue grew 14%.
Why the mix effect is hard to size
Duolingo does not report gross margin by revenue line in the material reviewed here. So the size of the mix effect cannot be calculated from public data. What can be said is that the filing's own explanation, a rise in subscription gross margin, describes a within-line improvement rather than a shift between lines. The 10-Q's wording therefore points to cost per subscriber falling, not to the mix doing most of the work.
The February letter offers a clue on direction. It said the 2025 gross margin gain was driven by rising subscription margins and a greater mix of revenue from subscriptions. That pairing suggests management treats the subscription line as the higher-margin one. A growing share therefore flatters the blended number at least slightly.
Why the Q4 gross margin guide falls 90 basis points
The Q3 letter guides Q4 gross margin to 72.0%, down 90 basis points sequentially as the company expands the rollout of AI-based features. Munson said on the call that Duolingo would put Video Call out to all paid users by the end of the year. She said the guide includes the expected transition away from Max.
Run the guide through the revenue forecast and the implied cost is visible. Q4 revenue of $322 million at a 72.0% margin implies cost of revenues near $90.2 million, about $6.3 million more than Q3. Revenue is guided up about $13 million. The implied incremental gross margin on that sequential revenue is roughly 52%, against about 75.7% year over year in Q3.
The Max sunset is a pricing decision dressed as a product one
Duolingo said it will stop selling new Max subscriptions and move existing Max individual subscribers to a Super family plan at the same price. Munson described Max as single digits of both paid subscribers and bookings. The company expects the bookings effect to be minimal, and it is still considering options for legacy Max family subscribers in 2027.
I think this is the most important strategic signal in the quarter. A premium AI tier existed because inference was expensive. Once inference became nearly free per call, the tier lost its reason to exist. Duolingo chose reach over price discrimination. The same pressure is visible in why cheaper tokens still produce rising AI bills. Unit prices fall, then usage and access expand to meet them.
Management also signalled on the call that it expects to spend some inference cost on free users next year. The CEO said most inference currently runs through outside inference providers, with some internal testing. Free users carry no payment processing fee and no subscription revenue. AI features aimed at them land almost entirely as cost. That is where the 2027 margin argument will be settled, and the company has deferred 2027 guidance to February.
Where this argument is weakest
The analysis above leans on a small number of disclosures, and several of them are softer than they look. Here is where a sceptic should push.
The cost per call is a management statement, not a disclosure
The $0.30 and sub-cent figures were spoken on an earnings call. They do not appear in the 10-Q or the shareholder letter. Duolingo has not defined what a call includes, how long an average call runs, or whether the figure covers speech recognition, voice synthesis and hosting or only model inference. A per-call number without call volume cannot be turned into a total cost. That is why the chart in this post is explicitly illustrative.
The 10-Q attribution is also qualitative. The filing says the margin gain was primarily attributable to lower feature based AI costs, but it does not say by how much. A 40 basis point annual gain is small. Hosting, payment fee rates or a shift in where subscribers buy could plausibly account for a meaningful share of it. My separation of drivers rests on the sequence of guidance and statements, not on line-item data. Treat it as an informed reading rather than a measurement.
The strongest counter-case
The best argument against this post is that the AI savings were in fact large, and Duolingo simply chose to show them in engagement rather than margin. On that view, words spoken per active user roughly doubling over the past year, as the CEO said on the call, is the return, and margin is the wrong scoreboard. I find that persuasive for the product. It does not change the financial point. Investors looking for AI savings in the gross margin line found 40 basis points, and the operating margin fell.
Meanwhile the bottom line moved the other way. Adjusted EBITDA margin was 25.5% in Q3 2026, down from 29.5%, and free cash flow fell 3% to $75.3 million. Research and development expense rose from $82.7 million to $107.7 million, about 30%. The quarter also carried an $11.4 million performance bonus accrual. Whatever the AI savings were worth at the gross margin line, operating costs absorbed more than all of it.
What this means for your AI feature cost per user
If you run a subscription product with AI features, Duolingo's year is a usable template, mostly because of what it reveals about sequencing. The order of decisions mattered more than the model choice.
Price the feature at today's cost, plan the tier for next year's
Duolingo launched Video Call in its top tier because $0.30 a call was unaffordable at scale. Within roughly two years the same feature cost under a cent and moved down a tier. If your premium tier exists mainly to recover inference cost, assume its rationale has a shelf life. Design the migration path before you need it. The per-feature accounting behind that decision is set out in the contribution margin method for AI features.
Measure per paying user, not per active user
Duolingo's cost per DAU fell about 13% while cost per subscriber fell about 2%. If you report AI efficiency per active user and most of your actives are free, you will overstate your progress. Track cost of revenues per paying customer, split into payment fees, hosting and AI, and you will see the saving that actually reaches the margin.
Test open-weight models on the high-volume feature first
The CEO's claim is that open-weight models are about as good as frontier models for the majority of users. That claim is strongest for a narrow, high-volume, conversational task like Video Call. I would test the swap on your single highest-volume AI feature, with a quality bar set in advance. Do that before touching anything where a wrong answer has consequences. The cost-per-task evidence for smaller models is collected in the cost per task comparison for small language models.
Frequently asked questions
What is Duolingo's gross margin?
Duolingo's gross margin was 72.9% in the third quarter of 2026, up from 72.5% in the same quarter of 2025, according to its 10-Q. For the first nine months of 2026 it was 72.8%, against 72.0% a year earlier. Full-year 2025 gross margin was 72.2%. The company guided Q4 2026 gross margin to 72.0%, which implies a full-year 2026 figure near 72.6% on its own revenue guidance.
Why did Duolingo's gross margin increase in Q3 2026?
The 10-Q says the increase was primarily attributable to higher subscription gross margin, reflecting lower feature based AI costs. The shareholder letter adds favourable hosting cost trends. Management had guided Q3 margin to about 71.0%, so the actual beat its own plan by roughly 1.9 points. The CFO said AI costs had not been as high as the company expected when it set that guidance.
How much does a Duolingo Video Call cost to run?
On the Q3 2026 earnings call, CEO Luis von Ahn said Video Call inference now costs well under $0.01 per call, down from about $0.30 per call at launch. He credited the reduction mainly to moving to open-source or open-weight models. The figure is a spoken management statement. Duolingo has not disclosed call volumes, call length or exactly which costs the per-call figure includes.
Why is Duolingo sunsetting Duolingo Max?
Max was created partly because some AI features were very expensive to serve. Once Video Call became cheap enough to include in the Super tier, Max lost its main differentiator. Duolingo said it will stop selling new Max subscriptions and move existing Max individual subscribers to a Super family plan at the same price. The CFO said Max is single digits of paid subscribers and bookings.
Does Duolingo use open-weight models?
Yes, according to management. The CEO said the drop in Video Call cost came mainly from moving to open-source or open-weight models, and that for most users those models perform about as well as frontier models. He also said most inference currently runs through outside inference providers, with some internal hosting being tested. The 10-Q describes the related costs only as hosting fees and third-party AI costs.
Will Duolingo gross margin fall in Q4 2026?
Duolingo guided Q4 2026 gross margin to 72.0%, down 90 basis points from Q3, as it expands AI-based features such as Video Call to all paid subscribers by year end. The guidance also reflects the planned transition away from the Max tier. Even at 72.0%, the quarter would sit about 3 points above the roughly 69% the company planned for this period in February.
Where to start this week
Pull your own last two quarters of cost of revenues and divide it by paying customers, not active users. Then split that number into payment fees, hosting and model spend. If you cannot produce the third component, you cannot tell whether a cheaper model improved your margin or simply funded more usage. That is the exact question Duolingo's filing leaves open.
If one of your AI features sits in a premium tier because it was expensive at launch, re-price it at today's per-call cost before your next pricing review. Duolingo's Max decision shows the tier can outlive its reason within two years.
Related analysis
For the wider pattern of falling unit prices meeting rising usage, read how Chegg's subscriber base reacted to ChatGPT, a consumer learning business on the other side of the same shift.
References
- Duolingo, Inc., Form 10-Q for the quarter ended 30 September 2026, filed November 2026. Used for revenue, cost of revenues, gross margin, revenue lines, operating expenses, key metrics and the cost of revenues definition.
- Duolingo, Inc., Q3 2026 shareholder letter (Form 8-K exhibit), 3 November 2026. Used for the hosting attribution, Q4 guidance, Max sunset and adjusted EBITDA figures.
- The Motley Fool, Duolingo (DUOL) Q3 2026 earnings call transcript, 4 November 2026. Used for the Video Call cost per call, open-weight models, Max mix and CFO comments.
- Duolingo, Inc., Q2 2026 shareholder letter, August 2026. Used for Q2 gross margin, the 71.0% Q3 guide and the 71.6% full-year expectation.
- Duolingo, Inc., Q4 2025 shareholder letter, February 2026. Used for Q4 2025 and full-year 2025 gross margin, the original 2026 margin plan and the Video Call move to Super.
- Duolingo, Video Call product page, accessed November 2026. Used for the feature description.
Weakest point in the source base: the per-call cost figures come from an earnings call transcript, not from a filing, and Duolingo discloses no AI cost total. Q1 2026 gross margin of 73.0% is taken from the Q1 shareholder letter as quoted in search results and was not opened directly. All computed ratios are Zan Digital arithmetic on reported figures, retrieved 11 November 2026.
Related reading