Verified
Duolingo
Embedded- Education app
- 900+ employees
- AI-first since Apr 2025
- OpenAI, Anthropic, open-weight models
- 20,500 course units in Q1 2026
Duolingo deploys AI mainly inside its learning app: AI content tools helped it publish 20,500 course units in Q1 2026, against 1,800 a quarter in 2024, a model called Birdbrain picks each learner's exercises, and Video Call gives subscribers an AI conversation partner. CEO Luis von Ahn declared Duolingo AI-first in Apr 2025, tying AI use to hiring and performance reviews. Duolingo uses OpenAI and Anthropic models internally and moved Video Call mostly to open-source models, cutting its cost to under $0.01 a call by Aug 2026.
duolingo.comTheir approach to AI
- The platform
- Duolingo builds its own AI layer: the Birdbrain personalization model, prompt templates for course content, an internal LLM Gateway that tracks model cost, and an agent platform on Temporal for engineers (Aug 2026).
- The models
- Duolingo launched Max on OpenAI's GPT-4 in Mar 2023, still uses OpenAI and Anthropic models internally, runs Video Call mainly on open-source models, and uses only Chinese models in China (Aug 2026).
- The reach
- Duolingo's machine learning sets exercise difficulty for every learner in an app with 58.7 million daily active users (Q2 2026), Video Call reaches Max and most new Super subscribers, and the Apr 2025 AI-first memo made AI use part of performance reviews.
- The catch
- Duolingo reports no firm-level return on AI; its letters tie gross margin to AI costs, which pulled 2025 margin down to 72.2% before lower per-unit costs lifted Q2 2026 to 72.6%, and an Oct 2026 investor suit alleges AI content hurt the product.
Snapshot
- Employees
- 900+
- Dec 31, 2025, including 430+ engineers; about 830 a year earlier
- Technology budget
- Not disclosed
- R&D expense $306.3M in 2025, up 30%, with GenAI costs up $3.6M within it
- AI/ML headcount
- Not disclosed
- 430+ engineers (Dec 2025); no separate AI or machine learning count
- Who owns AI
- Luis von Ahn, CEO
- Set the Apr 2025 AI-first policy; Andrew Hogue was VP and Head of AI by May 2026
- Platform
- Built in-house
- Birdbrain, content tools, LLM Gateway and an agent platform on Temporal
- Models
- OpenAI, Anthropic, open-weight
- Video Call mainly on open-source models; Chinese models in China (Aug 2026)
- Reach
- 58.7M daily learners
- Q2 2026; AI-built content and AI-chosen exercises run through the app
- Use cases
- Content, tutoring, testing, ads, code
- In production across the app, the Duolingo English Test and engineering
Key findings
Duolingo published 20,500 course units in Q1 2026, up from 7,100 a quarter in 2025 and 1,800 a quarter in 2024, and credits AI tools that change how quickly it can create content; the same approach let Duolingo add nearly 150 new language courses in Q1 2025.
Luis von Ahn's Apr 2025 AI-first memo said Duolingo would gradually stop using contractors for work AI can handle and add headcount only where a team cannot automate more; von Ahn said weeks later that Duolingo was still hiring at the same speed, and headcount rose from about 830 at the end of 2024 to more than 900 at the end of 2025.
Luis von Ahn said on Aug 5, 2026 that Video Call, once expected to cost about $0.30 a call, now costs under $0.01, mainly because Duolingo moved to open-source models, while Duolingo still uses OpenAI and Anthropic models internally and must use Chinese models in China.
Duolingo's AI costs move its gross margin: AI costs for features like Video Call helped pull 2025 gross margin down to 72.2% from 72.8%, lower per-unit AI costs lifted Q1 2026 gross margin 190 basis points to 73.0%, and CFO Gillian Munson put AI in cost of revenue in the tens of millions of dollars in Aug 2026.
Duolingo engineers run coding agents on an internal platform built on Temporal with Claude Agents SDK, Codex CLI and OpenAI Agents SDK runtimes, the last routed through an internal LLM Gateway that tracks cost (Aug 2026), and an agent-driven pipeline merged 250 iOS unit test pull requests in about 17 weeks (Jun 2026).
Where Duolingo sits
- Stage
- Embedded
Duolingo sits at Embedded: AI content tools helped it publish 20,500 course units in Q1 2026 against 1,800 a quarter in 2024, machine learning has set exercise difficulty since Birdbrain in 2020, Video Call puts an AI conversation partner in its paid tiers, and an Apr 2025 AI-first memo made AI use part of hiring and performance reviews while phasing out contractors for work AI can handle. Duolingo reports AI's financial effect through gross margin every quarter and credits a machine learning ad model with tens of millions of dollars a year in incremental revenue (Mar 2025). Duolingo's case has limits: it has published no figure for how many of its 900+ employees use AI each week, and no firm-level AI return.
How the stages are defined - Ownership
- CEO Luis von Ahn set the AI-first policy; Andrew Hogue, VP and Head of AI by May 2026; no chief AI officer
- Platform
- Hybrid: Own models, gateway and agent platform; outside foundation models
- Models
- OpenAI and Anthropic, open-weight models, Chinese models in China
- Hosting
- AWS and Google Cloud host the app and data; OpenAI and Anthropic models used as outside services; where open-weight and China models run not disclosed
- Reach
- Machine learning in lessons for 58.7M daily users (Q2 2026)
How they deploy
AI-first, set by the CEO
Luis von Ahn told Duolingo staff in an Apr 2025 all-hands email, shared on Duolingo's LinkedIn page, that Duolingo would be AI-first: it would gradually stop using contractors for work AI can handle, look for AI proficiency in hiring, evaluate AI use in performance reviews, and add headcount only when a team cannot automate more of its work. Von Ahn wrote weeks later on LinkedIn that he does not see AI replacing what Duolingo's employees do and that Duolingo was continuing to hire at the same speed as before (Entrepreneur, May 2025).
Teach better with AI, even at a cost to margin
Luis von Ahn wrote in Duolingo's Q3 2025 letter that his conviction that AI can change how Duolingo teaches drove a shift toward teaching better and user growth over monetization. Duolingo's Feb 2026 plan set out to move Video Call into the Super Duolingo tier, which Duolingo said would give about ten times as many learners access, and to add AI-powered speaking practice for free and paying learners. Duolingo guided gross margin down to about 69% for the rest of 2026 to pay for that wider access, then raised its full-year outlook to about 71.6% in Aug 2026 on AI cost trends, and Luis von Ahn said Duolingo may sunset its Max tier, with a decision due within a couple of quarters.
Build the AI layer, swap the models
Duolingo builds its own AI layer, including Birdbrain, prompt templates for course content, Video Call's prompt and memory design, an LLM Gateway and an agent platform, and buys foundation models. Duolingo launched Max on OpenAI's GPT-4 in Mar 2023. Luis von Ahn said in Aug 2026 that open-source models trail the frontier labs by perhaps three to six months, that many uses do not need the smartest model, and that Duolingo will keep a mix of models with a growing open-weight share.
Content at scale, with human review
Duolingo's 2025 Form 10-K says AI lets it generate high-quality course content at materially faster speeds. Duolingo's learning designers fill in prompt templates and pick and edit the model's output, with teaching experts keeping the final say (Jun 2023), and Luis von Ahn said in May 2026 that AI content passes AI and human evaluations and small tests with learners before wider release, because careless use of AI produces a lot of slop.
Agents for engineering work
Duolingo's engineers shipped their first coding agent, a feature flag remover built on OpenAI's Codex CLI and Temporal, in Dec 2025, then a form-based way for staff, including product managers and researchers, to launch agentic workflows without writing code. By Aug 2026 Duolingo had an agent platform on which a developer defines an agent once and runs it from Slack, internal sites, a command line or other workflows, which Duolingo says cut the time to build a production-ready agent from several weeks to about 10 minutes.
- Deployment model
- Duolingo builds AI into one shared technology platform that serves the Duolingo app, its Math, Music and Chess courses and the Duolingo English Test, and had more than 430 engineers among more than 900 employees at the end of 2025. Duolingo's learning designers and content developers write the prompts and review AI output for courses, while engineering teams run Video Call, the Birdbrain personalization model, the ad decision model, the LLM Gateway and the agent platform. Duolingo has no chief AI officer among its executive officers: CEO Luis von Ahn set the AI-first policy, co-founder Severin Hacker was CTO in the Apr 2026 proxy statement, Natalie Glance leads engineering as Chief Engineering Officer, and Andrew Hogue was VP and Head of AI by May 2026.
- Model providers
- Duolingo used OpenAI's GPT-3 to draft Duolingo English Test items from 2022 and launched Duolingo Max on OpenAI's GPT-4 in Mar 2023. Luis von Ahn said on Aug 5, 2026 that Duolingo still uses many proprietary models from OpenAI and Anthropic internally, has moved Video Call mainly to open-source models, would move everything to open-weight models if it could, and uses only Chinese models in China because U.S. models cannot be used there. Duolingo's coding agents run on Anthropic's Claude Code and Claude Agents SDK and on OpenAI's Codex CLI and Agents SDK, and the example agent in Duolingo's Aug 2026 platform post names OpenAI's gpt-5.5. Duolingo has not named its open-weight models or its providers in China.
- Data and infrastructure
- Duolingo's 2025 Form 10-K says AWS and Google Cloud host a significant portion, if not almost all, of its product traffic, data storage and data processing, and that learners complete nearly 2 billion exercises a day, which Duolingo calls the world's largest learning dataset. Duolingo's first coding agent runs on a separate AWS ECS instance and its iOS test pipeline stores state in AWS S3. Duolingo's agent platform routes OpenAI Agents SDK calls through an internal LLM Gateway that tracks cost and usage and abstracts the model provider (Aug 2026), and its engineers extended cloud cost tracking to OpenAI in 2024. Duolingo has not said where its open-weight models or its models in China run.
Platforms and internal tools
Birdbrain2020
Birdbrain is Duolingo's machine learning model that estimates whether a learner will answer a given exercise correctly and uses that estimate to build lessons at the right difficulty. Duolingo said Birdbrain personalized more than 20% of lessons by Oct 2020, and Duolingo's 2025 Form 10-K describes machine learning that evaluates every learner's answer to every exercise each day.
Duolingo MaxMar 2023
Duolingo Max is the subscription tier above Super Duolingo, launched in Mar 2023 on OpenAI's GPT-4, with AI features that Duolingo's 10-K names as Roleplay, Explain My Answer and Video Call; Duolingo's Max announcement post, since updated, lists it in 188 countries on iOS and Android. Duolingo began moving Video Call down to Super Duolingo in 2026, and Luis von Ahn said in Aug 2026 that Duolingo may sunset Max and would decide within the next couple of quarters.
- Users
- 5% of paid subscribers (Dec 2024)
- Models
- OpenAI GPT-4 at launch
Video Call2024
Video Call lets learners hold spoken conversations with Lily, a Duolingo character, and beginners with Falstaff. Learning designers write the system instructions, the model keeps a list of facts about each learner between calls, and mid-call checks end the call if a learner raises something inappropriate. Duolingo said in Aug 2026 that most new Super Duolingo subscribers have access and that existing Super subscribers would follow later in 2026.
- Users
- Max and most new Super subscribers (Aug 2026)
- Models
- Mainly open-source models (Aug 2026)
AI course content toolsBy 2023
Duolingo's content tools let learning designers fill prompt templates with the language, level, theme and grammar focus, generate candidate exercises with a large language model, and keep and edit the best ones. Duolingo used AI-generated content to add nearly 150 new language courses in Q1 2025 and to take the courses for its nine most-learned languages up to CEFR B2 by Apr 2026.
LLM GatewayBy Aug 2026
Duolingo's LLM Gateway is an internal service between its software and model providers that tracks cost and usage and abstracts the provider, so agents on Duolingo's agent platform can switch providers behind one interface.
Agent platformAug 2026
Duolingo's agent platform pairs an agent registry, where a developer sets an agent's system prompt, MCP servers, repositories, model and output type, with a Temporal workflow called AgentWorkflow that runs the agent and returns its output. Agents on the platform fix CI failures, address code review comments and back a Slack bot for release managers, and an evaluation suite runs each agent against written test scenarios and grades its output and code changes.
- Models
- Claude Agents SDK, Codex CLI and OpenAI Agents SDK runtimes
Agentic workflowsDec 2025
Agentic workflows let Duolingo staff launch a coding agent from a form with a prompt, a target repository and optional parameters, using a shared CodingAgent library that wraps Codex CLI and the Claude Code SDK. The agents remove deprecated feature flags, start or stop experiments, edit Terraform and open pull requests from a bot account.
- Models
- OpenAI Codex CLI, Anthropic Claude Code SDK
Ad decision modelBy Mar 2025
Duolingo's ad decision model, an XGBoost model later turned into a contextual bandit, chooses whether a free learner sees an in-house ad for a subscription or a paid network ad after each lesson, replacing rule-based logic.
Vendors and partners
OpenAI2022
GPT-4 behind Duolingo Max at launch; GPT-3 for English test items; models and Codex tools used internally · Duolingo's English test team used GPT-3 to draft test content from 2022, Max launched on GPT-4 in Mar 2023, and Luis von Ahn said in Aug 2026 that Duolingo will keep using OpenAI models internally.
AnthropicBy Dec 2025
Models used internally; Claude Code and Claude Agents SDK in Duolingo's coding agents · Duolingo's agentic workflows wrap the Claude Code SDK (Dec 2025), its iOS test pipeline runs Claude Code (Jun 2026), and Luis von Ahn named Anthropic among the proprietary providers Duolingo uses (Aug 2026).
Amazon Web Services
Cloud host for product traffic and data; ECS sandboxes for agents; S3 for the iOS test pipeline · Duolingo engineers cut cloud spending by about 20% on an annualized basis in a few months in 2024 and extended cost tracking beyond AWS to OpenAI.
Google Cloud
Cloud host for product traffic and data alongside AWS · Duolingo's 2025 Form 10-K names AWS and Google Cloud as the hosts of a significant portion, if not almost all, of its product traffic.
TemporalBy Dec 2025
Workflow engine under Duolingo's coding agents and agent platform
Leadership
AI, data and platform leaders
Executives
AI org map
Founder-led; AI-first policy set by the CEO; AI built by engineering and learning teams on one shared platform
Duolingo has no chief AI officer among the executive officers in its Apr 2026 proxy statement. CEO Luis von Ahn set the AI-first policy in Apr 2025, co-founder Severin Hacker was CTO in that proxy statement, Natalie Glance runs engineering as Chief Engineering Officer, Andrew Hogue was VP and Head of AI by May 2026, and CFO Gillian Munson reports AI costs to investors. Duolingo has not published reporting lines below the CEO, so the lines between units are inferred.
Duolingo had more than 900 employees, including more than 430 engineers, at Dec 31, 2025, up from about 830 employees and 380 engineers a year earlier, and does not publish a separate count of AI or machine learning staff. Duolingo's AI work spans engineers who build the agent platform and developer tooling, machine learning engineers behind Birdbrain and the ad decision model, and learning designers and content developers who write prompts and review AI-generated lessons. Duolingo's research site lists an AI and machine learning group beside learning, assessment and data science teams, but its publication list stops in 2021. Andrew Hogue was VP and Head of AI by May 2026; Duolingo has not published the size of his team.
Reporting lines
CEO and Chairman
Chief Technology Officer
Co-founder; title per the Apr 2026 proxy statement; reporting line inferred
Engineering
Natalie Glance, Chief Engineering Officer
430+ engineers (Dec 2025); reporting line inferred
Agent platform and developer tooling
Agent registry, AgentWorkflow on Temporal and the LLM Gateway (Aug 2026); team lead not named
Machine learning
Birdbrain, the ad decision model and chess models; placement inferred
AI
Andrew Hogue, VP and Head of AI
In the role by May 2026; reporting line and team size not published
Learning and curriculum
Learning designers and content developers who write prompts and review AI course content; reporting line not published
Duolingo English Test
Uses AI to help produce test content; reporting line not published
Finance
Gillian Munson, CFO
Reports AI costs and their effect on gross margin each quarter
Duolingo's 2026 proxy statement lists the CEO, CTO, CFO, Chief Engineering Officer, Chief Business Officer and General Counsel as executive officers and no AI-specific officer. Former CFO Matt Skaruppa advises Duolingo through Nov 20, 2026. Duolingo's board risk oversight sits with its Audit, Risk and Compliance Committee, whose remit names information technology, privacy and cybersecurity but not AI.
Where the AI workforce sits
Engineering, including machine learning and the agent platform
Estimate from the 2025 Form 10-K: more than 430 engineers among more than 900 employees. Duolingo does not publish how many of them work on AI.
~48%Learning design, product, design and business teams
Estimate: the remaining staff. Learning designers write and review AI-generated content, and the Apr 2025 memo said most functions would change through specific AI initiatives.
~52%
Named people
Use cases
Duolingo runs AI in production across course content creation, lesson personalization, conversation practice (Video Call, Roleplay, Explain My Answer), the Duolingo English Test, ad decisions, chess coaching and software engineering. Duolingo has published a business return for one: the machine learning model that picks which ad to show added millions of dollars a year in incremental revenue within months, and tens of millions after refinements and bug fixes (Mar 2025). Duolingo credits AI tools for raising course output from 1,800 units a quarter in 2024 to 20,500 in Q1 2026, an output measure with no cost or revenue attached. Duolingo reports no firm-level AI return; its quarterly letters instead tie gross margin to AI costs, which rose with Max in 2025 and fell per unit in 2026.
| Use case | Area | Status | ROI |
|---|---|---|---|
| AI course content creation Duolingo's learning designers generate exercises and lessons with large language models from prompt templates and review the output; Duolingo used AI-generated content to add nearly 150 new language courses in Q1 2025, to extend its nine most-learned languages to CEFR B2 by Apr 2026, and to publish 20,500 course units in Q1 2026 against 1,800 a quarter in 2024. | Learning content | In production | |
| Birdbrain lesson personalization Birdbrain predicts whether each learner will get each exercise right and builds lessons at the right difficulty from that estimate. | Learning product | In production | |
| Video Call Video Call lets learners practice speaking with Lily or, for beginners, Falstaff; Duolingo said in May 2026 that average words spoken per learner in Video Call more than doubled over the prior year, and Duolingo began opening the feature to Super Duolingo subscribers in 2026. | Learning product | In production | |
| Explain My Answer Explain My Answer gives learners personalized feedback on grammar and vocabulary mistakes; it began in Duolingo Max and became free for most learners in seven courses in Jan 2026. | Learning product | In production | |
| Roleplay Roleplay lets Duolingo Max subscribers practice conversations with Duolingo characters; Duolingo cut its model costs significantly through optimizations in Q1 2025. | Learning product | In production | |
| Spoken answers and Speaking Adventures Duolingo lets free and paying learners speak their answers to most exercises and added Speaking Adventures, a free AI-powered lesson type built around real-world spoken tasks (announced Feb 2026, live by May 2026). | Learning product | In production | |
| Test content generation The Duolingo English Test team used OpenAI's GPT-3 from at least 2022 to draft passages and answer choices that test developers review for accuracy, fairness and bias, and Duolingo's 2025 Form 10-K says AI still helps produce test content. | Duolingo English Test | In production | |
| Ad decision model A machine learning model chooses whether each free learner sees an in-house subscription ad or a paid network ad after a lesson. | Monetization | In production | Incremental revenue of millions a year at launch, tens of millions after refinements and bug fixes (Mar 2025) |
| Chess Game Review and human-like chess models Duolingo trains a family of human-like neural networks, set by rating strength, to estimate win, draw and loss chances and pick out the key moments in a learner's chess game. | Chess | In production | |
| Coding agents and agentic workflows Duolingo's coding agents remove feature flags, fix CI failures, address code review comments, help upgrade JVM services and open pull requests, and run from Slack, internal sites and a command line on Duolingo's agent platform. | Engineering | In production | |
| AI-generated iOS unit tests A Temporal pipeline runs Claude Code to write iOS unit tests, opens draft pull requests and retries CI fixes; 250 pull requests merged in about 17 weeks, each approved by a human engineer, and MVVM test coverage rose from 9% to 30% (Jun 2026). | Engineering | In production |
Adoption
Duolingo spread AI through its learning product first: Birdbrain personalization from 2020, GPT-3 drafting of English test items from 2022, the GPT-4 Max tier from Mar 2023 and Video Call from 2024, then AI course content at scale from 2024 to 2026. Duolingo's AI-built content and AI-chosen exercises sit in an app with 58.7 million daily and 140.6 million monthly active users (Q2 2026); Explain My Answer became free for most learners in seven courses in Jan 2026, and most new Super subscribers had Video Call by Aug 2026. Inside Duolingo, the Apr 2025 memo made AI use part of performance reviews and engineers moved to coding agents from Dec 2025, but Duolingo has published no figure for how many employees use AI tools each week.
By the numbers
Rollout, first pilot to today
Oct 2020
Duolingo introduced Birdbrain, a machine learning model that predicts how each learner will do on each exercise, and said it was personalizing more than 20% of lessons.
Apr 2022
Duolingo said its English test team uses OpenAI's GPT-3 to draft test passages and answer choices, with human reviewers in the loop.
Mar 2023
Duolingo launched Duolingo Max, a subscription tier above Super Duolingo built on OpenAI's GPT-4.
Jun 2023
Duolingo described how its learning designers write lesson exercises with a large language model from prompt templates and edit the output.
Jan 2024
Duolingo confirmed it cut about 10% of its contractor workforce at the end of 2023, with part of the reason attributed to AI; a spokesperson said GPT translates sentences and human experts validate the output.
Nov 2024
Duolingo reported launching Video Call inside Duolingo Max in Q3 2024, with Max available to about half of its daily active users.
Feb 2025
Duolingo said Max reached 5% of paid subscribers by the end of 2024 and that AI and automation tools had significantly scaled lesson content creation, driving efficiency gains in R&D.
Apr 2025
Luis von Ahn declared Duolingo AI-first in an all-hands memo shared on Duolingo's LinkedIn page, phasing out contractors for work AI can handle and tying AI use to hiring and performance reviews.
May 2025
Duolingo reported adding nearly 150 new language courses in Q1 2025 with AI-generated content and cutting Roleplay's model costs ahead of schedule.
Dec 2025
Duolingo's engineers shipped their first coding agent, a feature flag remover built on OpenAI's Codex CLI and Temporal, followed by form-based agentic workflows for any employee.
Feb 2026
Duolingo set a plan to move Video Call into Super Duolingo and add AI-powered speaking practice for free learners, guiding to a lower gross margin to widen access to AI features.
May 2026
Duolingo published 20,500 course units in Q1 2026, more than ten times its 2024 quarterly pace, crediting its AI content tools.
Aug 2026
Duolingo described an agent platform behind an internal LLM Gateway, and Luis von Ahn said Video Call had moved mainly to open-source models.
Economics & controls
Spend and ROI
Technology spending
Duolingo does not publish a technology budget. Duolingo's research and development expense rose 30% to $306.3 million in 2025, including $8.8 million more for web services and technology, of which $3.6 million was higher GenAI cost, and its cost of revenue of $288.1 million includes hosting fees and AI costs. Duolingo had $44.4 million of purchase commitments for hosting, web services and AI at Dec 31, 2025, according to its 2025 Form 10-K. Duolingo's research and development expense rose 22% to $175.2 million in the first half of 2026, including $5.5 million more for software and third-party AI, according to its Q2 2026 Form 10-Q.
What AI itself costs
Duolingo's 2025 gross margin fell to 72.2% from 72.8%, which Duolingo attributed to AI costs for features like Video Call and a shift toward lower-margin advertising. CFO Gillian Munson said on Aug 5, 2026 that AI in cost of revenue runs in the tens of millions of dollars and internal AI use is closer to $10 million, without naming a period, after saying in May 2026 that operating expenses carried pretty big increases in internal AI costs. Luis von Ahn said Video Call, expected to cost about $0.30 a call at launch, now costs under $0.01, mainly because of open-source models.
Reported ROI
Duolingo publishes no firm-level AI return. Duolingo's one use-case return is in a Mar 2025 engineering post, which credits its machine learning ad decision model with millions of dollars a year in incremental revenue within months of launch, rising to tens of millions after the team refined the model and fixed older bugs. Duolingo credits AI content tools for raising course output from 1,800 units a quarter in 2024 to 20,500 in Q1 2026 but has not tied that output to a cost or revenue figure, and its Q4 2024 letter said AI and automation tools helped drive efficiency gains in research and development, without a figure.
Margins and AI cost trends
Duolingo's gross margin moves with AI costs: it fell 100 basis points year over year in Q2 2025 on Max's AI costs, less than the 300 basis points Duolingo expected, and rose 190 basis points to 73.0% in Q1 2026, driven mainly by lower per-unit AI costs. Duolingo guided to about 69% gross margin for most of 2026 to give more learners AI features, then in Aug 2026 raised its full-year outlook to about 71.6% based on AI cost trends.
Productivity markers
Luis von Ahn said in May 2026 that the number of A/B tests Duolingo can run per engineer is rising moderately, which he attributed to AI use in engineering and product, and in Aug 2026 that experiments have grown faster than headcount, so each person puts out more experiments every week; Duolingo gave no figure. Duolingo's engineering posts report that its agent platform cut the time to build a production-ready agent from several weeks to about 10 minutes (Aug 2026) and that AI-written iOS tests lifted MVVM coverage from 9% to 30% (Jun 2026).
Headcount
Duolingo had about 720 employees at the end of 2023, about 830 at the end of 2024 and more than 900 at the end of 2025. Duolingo confirmed in Jan 2024 that it had cut about 10% of its contractors at the end of 2023, partly because of AI, and Luis von Ahn's Apr 2025 memo said Duolingo would gradually stop using contractors for work AI can handle and add staff only where a team cannot automate more; von Ahn said in May 2025 that Duolingo was still hiring at the same speed.
Governance and risk
Human review of AI content
Duolingo keeps humans in charge of AI-generated teaching material: learning designers write the prompts and edit the output, with teaching experts having the final say (Jun 2023), and Duolingo English Test developers review GPT-drafted items for accuracy, fairness and bias (Apr 2022). Luis von Ahn said in May 2026 that AI content now passes AI and human evaluations and small tests with learners before wider release.
Guardrails in Video Call
Duolingo's Video Call runs mid-call evaluations that end the call if a learner raises something inappropriate and that prompt Lily to rephrase when a learner seems confused, and a closing step limits call length. Video Call stores a list of facts learned about each learner and feeds it into the next call's instructions (Apr 2025).
Controls on coding agents
Duolingo runs its first coding agent in full-auto mode on a separate AWS ECS instance, which its engineers call much safer for bypassing approvals, assigns each pull request to the requesting employee, and controls repository access through a shared GitHub App (Dec 2025). Every merged AI-written iOS test was approved by a human engineer (Jun 2026), and the Aug 2026 agent platform tests agents against written scenarios with deterministic graders that check the code diff, cap the number of changed files and restrict edits to set paths, and shows every tool call's inputs, outputs and retries in Temporal.
Cost tracking
Duolingo's internal LLM Gateway tracks cost and usage for the agent model calls routed through it and abstracts the provider (Aug 2026), and Duolingo extended its cloud cost tracking to OpenAI in 2024. CFO Gillian Munson said in May 2026 that AI costs rise when features launch and fall as teams optimize, and that Duolingo adds more AI as costs drop.
Disclosed risks and regulation
Duolingo's 2025 Form 10-K warns that generative AI features such as Duolingo Max can produce inaccurate or biased output, that copyright in AI output is uncertain, and that Duolingo depends on third-party AI models whose pricing and terms it cannot control. Duolingo's 10-K lists AI rules in Colorado, Texas, California and China, including China's labeling rules for AI-generated content, and Luis von Ahn said in Aug 2026 that Duolingo must use Chinese models in China. A proposed securities class action filed against Duolingo on Oct 7, 2026 alleges that Duolingo hid how rapidly generated AI content was degrading its product; the claims are unproven, and Duolingo had not responded in court as of Oct 2026.
Board oversight
Duolingo's board oversees risk through its Audit, Risk and Compliance Committee, whose remit covers information technology, privacy and cybersecurity, according to the Apr 2026 proxy statement; Duolingo has not disclosed a board committee or executive dedicated to AI risk.
Recent signals
Oct 7, 2026
A Michigan public pension fund filed a proposed securities class action in the U.S. District Court for the Western District of Pennsylvania against Duolingo, Luis von Ahn and former CFO Matt Skaruppa, alleging Duolingo hid that rapidly generated AI content, more ads and subscription upsells were hurting its product and user growth; the claims are unproven.
Sep 8, 2026
Duolingo described Chess Game Review, which uses human-like neural networks Duolingo trains itself, and said it hired Reid McIlroy-Young, creator of the Maia chess engine.
Aug 5, 2026
Luis von Ahn said Video Call now costs under $0.01 a call, mainly after a move to open-source models, that Duolingo keeps using OpenAI and Anthropic models internally and Chinese models in China, and that Duolingo may sunset Max; CFO Gillian Munson put AI in cost of revenue in the tens of millions of dollars.
Aug 5, 2026
Duolingo reported a 72.6% gross margin for Q2 2026, ahead of its 71% outlook because of a measured pace of AI feature expansion and AI cost efficiencies, raised its full-year gross margin outlook to about 71.6%, and said most new Super subscribers have Video Call.
Aug 4, 2026
Duolingo described its agent platform, which runs agents on Claude Agents SDK, Codex CLI or OpenAI Agents SDK, routes OpenAI Agents SDK calls through an internal LLM Gateway, and cut the time to build a production-ready agent from several weeks to about 10 minutes.
Jun 24, 2026
Duolingo reported that a Claude Code pipeline merged 250 iOS unit test pull requests in about 17 weeks, each approved by an engineer, lifting MVVM test coverage from 9% to 30%.
May 29, 2026
Andrew Hogue, Duolingo's VP and Head of AI and formerly director of AI engineering for Google Workspace, spoke on a WBUR Festival panel about how AI is changing jobs and business.
May 4, 2026
Duolingo published 20,500 course units in Q1 2026, crediting AI content tools, and reported gross margin up 190 basis points to 73.0%, driven mainly by lower per-unit AI costs.
May 4, 2026
Luis von Ahn said AI content at Duolingo passes AI and human evaluations and small tests with learners before release, and CFO Gillian Munson said internal AI costs had risen sharply.
Mar 30, 2026
Duolingo published research finding that learners who used Video Call with Lily or Falstaff improved their speaking more than learners who took only regular lessons.
Feb 26, 2026
Duolingo set a 2026 plan to move Video Call into Super Duolingo, reaching about ten times as many learners, and to add AI speaking practice for free learners, guiding gross margin down to about 69%.
Jan 1, 2026
Duolingo made Explain My Answer, a Duolingo Max AI feedback feature, free for most learners in its Spanish, French, German, Japanese, Portuguese, Italian and Korean courses.
Dec 17, 2025
Duolingo described an AI-powered workflow that upgrades its JVM services by combining OpenRewrite with AI agents that fix build errors in a loop, which sped the work up by weeks.
Dec 11, 2025
Duolingo opened agentic workflows to employees, including product managers and researchers, through forms that launch Codex CLI or Claude Code SDK agents without writing code.
Nov 5, 2025
Luis von Ahn said Duolingo would invest more in teaching better and favor user growth over monetization, citing his conviction that AI can change how Duolingo teaches, and Duolingo raised its 2025 gross margin outlook on lower-than-expected AI costs.
Quick answers
- How does Duolingo deploy AI?
- Duolingo deploys AI mainly inside its learning app: AI content tools helped it publish 20,500 course units in Q1 2026, against 1,800 a quarter in 2024, a model called Birdbrain picks each learner's exercises, and Video Call gives subscribers an AI conversation partner. CEO Luis von Ahn declared Duolingo AI-first in Apr 2025, tying AI use to hiring and performance reviews. Duolingo uses OpenAI and Anthropic models internally and moved Video Call mostly to open-source models, cutting its cost to under $0.01 a call by Aug 2026.
- What does Duolingo use AI for?
- Duolingo runs AI in production across course content creation, lesson personalization, conversation practice (Video Call, Roleplay, Explain My Answer), the Duolingo English Test, ad decisions, chess coaching and software engineering. Duolingo has published a business return for one: the machine learning model that picks which ad to show added millions of dollars a year in incremental revenue within months, and tens of millions after refinements and bug fixes (Mar 2025). Duolingo credits AI tools for raising course output from 1,800 units a quarter in 2024 to 20,500 in Q1 2026, an output measure with no cost or revenue attached. Duolingo reports no firm-level AI return; its quarterly letters instead tie gross margin to AI costs, which rose with Max in 2025 and fell per unit in 2026.
- Who leads AI at Duolingo?
- AI at Duolingo is led by Severin Hacker (Co-Founder, Chief Technology Officer and Director); Natalie Glance (Chief Engineering Officer); Andrew Hogue (VP, Head of AI). Ownership of AI on the org chart: CEO Luis von Ahn set the AI-first policy; Andrew Hogue, VP and Head of AI by May 2026; no chief AI officer.
- What AI platform does Duolingo use?
- Duolingo builds its own AI layer: the Birdbrain personalization model, prompt templates for course content, an internal LLM Gateway that tracks model cost, and an agent platform on Temporal for engineers (Aug 2026).
- Which AI models does Duolingo use?
- Duolingo launched Max on OpenAI's GPT-4 in Mar 2023, still uses OpenAI and Anthropic models internally, runs Video Call mainly on open-source models, and uses only Chinese models in China (Aug 2026).
- How many Duolingo employees use AI?
- Duolingo's machine learning sets exercise difficulty for every learner in an app with 58.7 million daily active users (Q2 2026), Video Call reaches Max and most new Super subscribers, and the Apr 2025 AI-first memo made AI use part of performance reviews.
- How far has AI spread at Duolingo?
- Duolingo sits at Embedded: AI content tools helped it publish 20,500 course units in Q1 2026 against 1,800 a quarter in 2024, machine learning has set exercise difficulty since Birdbrain in 2020, Video Call puts an AI conversation partner in its paid tiers, and an Apr 2025 AI-first memo made AI use part of hiring and performance reviews while phasing out contractors for work AI can handle. Duolingo reports AI's financial effect through gross margin every quarter and credits a machine learning ad model with tens of millions of dollars a year in incremental revenue (Mar 2025). Duolingo's case has limits: it has published no figure for how many of its 900+ employees use AI each week, and no firm-level AI return.
Open questions
- How many of Duolingo's more than 900 employees use AI tools each week, and how do performance reviews measure AI use under the Apr 2025 memo?
- How much revenue do Duolingo Max and Video Call bring in, given that Duolingo last reported Max as 5% of paid subscribers at the end of 2024?
- What period do the Aug 2026 figures cover, tens of millions of dollars of AI in cost of revenue and about $10 million of internal AI use, and how do they split by provider?
- Which open-weight models run Video Call, where are they hosted, and which Chinese models does Duolingo use in China?
- How many contractors has Duolingo stopped using since the Apr 2025 memo, and what did that save?
- What return do Video Call, Roleplay, Explain My Answer and Speaking Adventures deliver in retention or subscriptions, given that Duolingo has published speaking research but no business figure?
- What do Duolingo's coding agents save in engineering time or cost, beyond agent build time and test coverage?
- When did Andrew Hogue become Duolingo's VP and Head of AI, whom does he report to, and how large is the AI team, given that Duolingo has not announced the role and its research site stops in 2021?
- Does Severin Hacker still hold the CTO title, given that Duolingo's investor site appeared in Nov 2026 search results describing him as Builder/Member of Technical Staff and former CTO while no filing records a change?
- Which Duolingo course units were generated with AI and what did they cost to produce, given that Duolingo reports 20,500 units in Q1 2026 but no unit cost or headcount saving?
- How does Duolingo's board oversee AI risk, given that the Audit, Risk and Compliance Committee's remit names cybersecurity and privacy but not AI?
Sources
Duolingo
- Making production-ready agents the default: Duolingo's agent platform (Aug 4, 2026)
- Duolingo blog: AI-generated iOS unit tests (Jun 24, 2026)
- Duolingo blog: Chess Game Review engineering (Sep 8, 2026)
- Duolingo blog: Video Call research report (Mar 30, 2026)
- Duolingo blog: courses now teach advanced content up to B2 (Apr 22, 2026)
- Duolingo blog: Explain My Answer now free (Jan 1, 2026)
- Duolingo blog: automating Golden Path upgrades with AI (Dec 17, 2025)
- Duolingo blog: agentic workflows (Dec 11, 2025)
- Duolingo blog: building AI agents (Dec 4, 2025)
- Duolingo blog: the AI behind every Video Call with Lily (Apr 22, 2025)
- Duolingo blog: machine learning for ad decisions (Mar 18, 2025)
- Duolingo blog: reducing cloud spending (Oct 30, 2024)
- Duolingo blog: writing lessons with a large language model (Jun 22, 2023)
- Duolingo blog: introducing Duolingo Max (Mar 14, 2023)
- Duolingo blog: how Duolingo's experts work with AI (Sep 14, 2022)
- Duolingo blog: machine learning in Duolingo English Test creation (Apr 6, 2022)
- Duolingo blog: introducing Birdbrain (Oct 7, 2020)
- Duolingo Research: team and publications
Filings, letters and investor materials
- Q2 2026 shareholder letter (Aug 5, 2026)
- Q2 2026 Form 10-Q (Aug 6, 2026)
- Q1 2026 shareholder letter (May 4, 2026)
- 2026 proxy statement (Apr 17, 2026)
- 2025 Form 10-K (filed Feb 27, 2026)
- Q4 2025 shareholder letter (Feb 26, 2026)
- CFO transition press release (Jan 12, 2026)
- Q3 2025 shareholder letter (Nov 5, 2025)
- Q2 2025 shareholder letter (Aug 6, 2025)
- Q1 2025 shareholder letter (May 1, 2025)
- 2024 Form 10-K (filed Feb 28, 2025)
- Q4 2024 shareholder letter (Feb 27, 2025)
- Q3 2024 shareholder letter (Nov 6, 2024)
- 2023 Form 10-K (filed Feb 29, 2024)
- Q2 2026 earnings call transcript, Investing.com (Aug 5, 2026)
- Q1 2026 earnings call transcript, Investing.com (May 4, 2026)
News
- OpenClassActions: Duolingo securities class action, filed Oct 7, 2026 (Oct 9, 2026)
- Kirby McInerney: securities class action filed against Duolingo, via Webull (Oct 8, 2026)
- WBUR Festival 2026: Andrew Hogue, VP and Head of AI at Duolingo (speaker page)
- Entrepreneur: Duolingo CEO clarifies AI stance after backlash (May 23, 2025)
- Tech.co: Duolingo to replace some workers with AI, CEO reveals (Apr 29, 2025)
- Pure AI: Duolingo leans into AI (Apr 29, 2025)
- TechCrunch: Duolingo cuts 10% of its contractor workforce as it embraces AI (Jan 9, 2024)
Data confidence
- Confirmed
- Stated in a primary or reputable public source.
- Estimated
- Derived or approximated from public reporting.
- Unconfirmed
- Reported, but not yet corroborated.
Duolingo profile · Enterprise AI diffusion report by Imagine · facts checked against their sources on Nov 1, 2026
