Duolingo
Conversational genAI tutor inside a premium subscription
Duolingo Max, a premium genAI tier built on GPT-4 (Video Call, Roleplay), accounted for about 5% of subscribers in late 2024 and drove bookings outperformance (+42% in Q4 2024), according to the earnings call.
Key points
- Premium Max tier adding genAI features (Explain My Answer, Roleplay, Video Call).
- Built on OpenAI GPT-4 with an in-house integration layer.
- About 5% of subscribers in late 2024, bookings up 42% year over year.
- Evidence A, confirmed status.
Objective
Create a higher subscription tier (Max) above Super, selling genAI features the free version cannot offer, to raise revenue per user and bookings.
The deployment
Duolingo Max adds genAI features on top of the subscription: Explain My Answer (the AI explains why an answer is wrong), Roleplay (a written conversation with an AI character), and Video Call, a video call with an animated character (Lily) that holds a spoken conversation generated in real time. These features run on GPT-4. Max is offered as a premium tier to a subset of users.
Results Proof A
Figures from Duolingo's Q4 2024 earnings call (transcript), where management explicitly attributes part of the bookings outperformance to Duolingo Max subscriptions. Financial results from a listed company.
How it works
Documented architectureThe stack in detail
- llm OpenAI GPT-4 Model behind Explain My Answer, Roleplay, and Video Call; generates explanations and dialogue turns in real time.
- outil Couche d'integration in-house Prompting with lesson context and the learner's answer, guardrails, and rendering in the interface.
- infra Application Duolingo (palier Max) Packaging of the genAI features into a subscription tier above Super.
- outil Rendu voix et personnage anime (Video Call) Syncs the real-time generated spoken conversation with the animated character Lily.
How it runs, concretely
For ops teams-
1Trigger human
The learner launches Explain My Answer, Roleplay, or Video Call from the app.
-
2Generation AI
The prompt (lesson context, learner's answer) is sent to the LLM, which produces the explanation or dialogue turn.
-
3Rendering site_app
The text or voice is rendered in the interface; for Video Call, synced with the animated character.
-
4Economic steering data / finance team
The team tracks inference cost per use and compares it with Max revenue to hold the margin.
The perceived quality of the generated conversation and explanation, which justifies the price of the Max tier. If the LLM response is generic or wrong, the value of the upgrade collapses.
How your customers perceive this type of use
Sourced studiesUn ecart net separe les annonceurs des consommateurs : 77% des annonceurs voient l'IA positivement contre 38% des consommateurs (Yahoo/Publicis, 2024). Les mesures implicites confirment le rejet declare : en EEG, les pubs generees par IA produisent une activation memorielle plus faible que les pubs traditionnelles et sont decrites comme agacantes, ennuyeuses et confuses (NIQ, 2024). La disclosure a un effet ambivalent : elle augmente fortement la confiance quand elle est remarquee (Yahoo/Publicis), mais 27% des jeunes consommateurs disent faire moins confiance a une entreprise dont la pub est creee par IA (IAB, 2024).
Acceptance conditions
- Une disclosure visible : quand la mention IA est remarquee, la confiance globale envers l'entreprise augmente de 96% (Yahoo/Publicis 2024)
- Une qualite visuelle suffisante : les visuels IA de basse qualite augmentent l'effort cognitif et distraient du message (NIQ 2024)
Red lines
- Le contenu IA non declare puis identifie : 72% des consommateurs disent que l'IA rend l'authenticite difficile a etablir (Yahoo/Publicis 2024) et les marques utilisant des pubs IA sont plus souvent jugees inauthentiques ou non ethiques par les consommateurs que par les dirigeants (IAB 2024)
- Les mannequins et personnes generes par IA : 46% des consommateurs n'en veulent pas dans la publicite, l'inquietude premiere etant les standards de beaute irrealistes (Attest 2025)
Sources: Yahoo / Publicis Media (terrain Ebco) 2024 · IAB (avec Attest) 2024 · NIQ (NielsenIQ) 2024 · Attest 2025
How to replicate
Inference, not sourcedData prerequisites
- structured educational content
- usage context (current answer, level)
- guardrails on the LLM's responses
Org prerequisites
- a product team able to package a premium tier
- tracking of inference costs
- a quality evaluation process for genAI outputs
Possible stack
- OpenAI / Anthropic or an equivalent LLM
- an in-house prompt and guardrail layer
- integration with the existing app
The plan, step by step
- Step 1Isolate a feature with high perceived value (the contextualized explanation) and prototype it on the LLM.Deliverable: A working prototype on real learner cases
- Step 2Evaluate the reliability of the outputs and set the guardrails.Deliverable: A quality evaluation grid and an acceptable error rate reached
- Step 3Package the premium tier: pricing, paywall, upgrade journey.Deliverable: A premium offer ready with per-use inference cost tracking
- Step 4Launch a pilot on a subset of users and measure conversion and retention.Deliverable: A read on tier conversion and churn on the pilot
- Step 5Launch more broadly and steer the margin: Max revenue against inference cost.Deliverable: The tier in production with a margin dashboard
First step: Isolate a feature with high perceived value (the contextualized explanation) and verify that the LLM makes it reliable before turning it into a paid subscription argument.
Sources
- S1 Duolingo (DUOL) Q4 2024 Earnings Call Transcript (The Motley Fool) Primary archive pending
- S2 Earnings call transcript: Duolingo Q4 2024 revenue beats forecast (Investing.com) Secondary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.