Loading today’s stories
Build1 distinct publisher3 min readUpdated
Luis von Ahn says open-weight models cut per-call inference cost more than 30-fold. The number that set Duolingo's paywall a year ago is now off by an order of magnitude and a half.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Duolingo co-founder and CEO Luis von Ahn told investors on the company's August 5th earnings call that an AI-powered video call, originally estimated at roughly $0.30, now costs the company less than $0.01, a decline he attributed mostly to a shift toward open models [1][3]. That figure matters less as a cost line than as a verdict on a pricing decision: the $0.30 estimate is what put Video Call inside Duolingo Max, the highest-priced tier, and it is now wrong by more than a factor of 30 [2][4].
Read the second-order effect first. Every operator who drew a paywall in the last eighteen months drew it around an inference quote, and inference quotes have not held. Duolingo did not get cheaper by negotiating; it got cheaper by swapping model classes for a task that did not need the expensive one. Von Ahn said open models can trail frontier laboratories by three to six months in some areas, but that a learner with a vocabulary of roughly 100 words does not require the most capable reasoning model available, and he reported no corresponding drop in quality for the task [3][6]. Duolingo still uses OpenAI and Anthropic models where their performance is required, and von Ahn said he expects the mix to move further toward open weights [7].
The per-call number is not a full accounting. Costs vary with conversation length, speech processing, animation, hosting and model choice, so $0.01 is a marginal figure and not a unit cost [5]. The aggregate is more instructive: CFO Gillian Munson said AI expenses inside cost of revenue run to tens of millions of dollars, with internal AI usage closer to $10 million [8]. Duolingo raised its full-year adjusted EBITDA margin forecast to 26.5% from the 25% it guided at the start of 2026, with AI savings contributing [9]. That is 1.5 points of margin partly financed by a model swap [15]. In the quarter itself, adjusted EBITDA of $77.3 million on revenue of $298.5 million works out to about 25.9%, and EBITDA fell 2% year over year while revenue rose 18% as the company spent on growth and wider product access [13][16].
Now the awkward part. According to runtimewire's account, Video Call was one of the main reasons to pay for Max, so moving it into Super removes the top tier's defining feature [10]. Most new Super subscribers already had access by the end of the second quarter, per the August 5th shareholder letter, with existing Super subscribers to follow later in 2026 [11]. Von Ahn said the company is weighing a call cap in Super with unlimited access in Max, adding a different expensive AI feature to Max, or retiring the tier outright [12].
The scale explains the temptation. Duolingo reported 58.7 million daily active users, up 23%, and 12.7 million paid subscribers, up 17% [13]. Roughly 22% of daily actives pay for anything [17]. Cheap inference lets the company put speaking practice in front of the other 78%, at the cost of the artifact it was using to sell the premium plan.
Watch which of the three options Duolingo picks, because it reveals whether Max was a product or a price fence. Watch whether a call cap appears in Super, since metering is the tell that marginal cost still bites at volume. And watch the 26.5% margin guide: if it holds while Video Call reaches existing Super subscribers, the savings were real rather than timing.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Duolingo co-founder and CEO Luis von Ahn said on the company's August 5th earnings call that Duolingo had reduced the cost of an AI-powered video call from roughly $0.30 to less than $0.01.
When Duolingo first introduced Video Call, the team estimated each interaction would cost about $0.30, and that expense helped determine the product's original placement inside Max, Duolingo's highest-priced subscription tier.
"It is now under $0.01 per video call," von Ahn said during the second-quarter earnings call, attributing most of the decline to a shift toward open models and adding that Duolingo had not observed a corresponding drop in quality for the task.
The reduction is greater than 30-fold, based on the two figures supplied by Duolingo.
The per-call measurement does not provide a complete view of Video Call's economics because costs can vary with conversation length, speech processing, animation, hosting and the model used.
Von Ahn said open models can trail systems from frontier laboratories by three to six months in some areas, but that the gap often has little practical effect on beginner conversation practice, and that a learner with a vocabulary of roughly 100 words does not require the most capable reasoning model available.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single-source, company-narrated with traceable primary statements
Every substantive figure originates with Duolingo management on a dated earnings call and shareholder letter, which makes the numbers attributable and specific, but the cluster contains exactly one publisher and no independent verification. The cost reduction, the quality-parity assertion and the educational efficacy findings are all company-produced; the article itself notes the per-call metric excludes speech processing, animation, hosting and model variation, and that the efficacy studies came from Duolingo rather than an independent evaluator.
Shipped in production at consumer scale, mid-rollout
This is not a demo: the open-model substitution is already serving a live consumer feature inside a business reporting 58.7 million daily active users and 12.7 million paid subscribers, and Video Call has reached most new Super subscribers with extension to existing Super subscribers still pending. Adoption is scored below the top band because the rollout is incomplete, call volumes are undisclosed, and the evidence covers one company's deployment rather than any broader pattern of open-weight substitution.
Mildly overstated: real cost move, softer than the headline framing
The cost reduction and the guidance raise are concrete and dated, so the story is not empty. The gap comes from framing: a partial per-call metric is used to carry an order-of-magnitude narrative, quality parity rests on assertion rather than published evaluation, and the 'took a pricing tier with it' framing describes options von Ahn is considering — including retiring Max — rather than any decision made. The article's own caveats keep the gap small rather than large.
Management-sourced on an earnings call, with in-house efficacy support
The disclosures were made by a CEO and CFO on a quarterly earnings call, where demonstrating margin expansion and defensible AI spend directly serves the company's interest, and where lowered cost is offered as justification for widening a paid feature's access while adjusted EBITDA declined 2%. The supporting educational evidence was produced by Duolingo's own research lab. No vendor payment, sponsorship or third-party funding relationship is disclosed in the source, so the scoring reflects issuer self-interest rather than any hidden conflict.
Moderate: specific primary quotes, no corroborating publisher
Confidence is supported by the specificity and traceability of the material — a named call date, quoted figures, a shareholder letter and reported quarterly results — but capped by a one-publisher cluster with zero independent corroboration or contradicting evidence, and by unquantified elements such as call volume, model identities and the eventual pricing decision.
leadership
Disney swaps raises for discounted stock and a full health-plan re-enrollment1 distinct publisher
science
Text watermarks land on 2 December. The detection they imply does not.1 distinct publisher
product
A 2x LLM bill is not a bug report: token spend is an observability problem1 distinct publisher
invest
Behind-the-meter gas is the data center buildout's real cost: 318 Mt a year1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 15, 2026