How Much Does It Cost to Build an App LikeDuolingo in 2026?

A lean first version of an app like Duolingo comes to about 1,320 to 2,080 hours of senior engineering in my estimate, which at $150 to $225 an hour is $198,000 to $468,000. Then you need a course to put inside it, and in my scenario one beginner course of 50 lessons adds another 200 to 400 hours.

A fuller launch, with adaptive review, AI explanations, pronunciation scoring, friends, a family plan and a web app, lands at about 2,160 to 3,490 engineering hours, or $324,000 to $785,250.

Those are my scenario estimates. They are not quotes and they are not measurements of what Duolingo spent. Duolingo is a trademark of Duolingo, Inc., and nothing in this article implies any relationship with the company.

Consider a founder who wants "Duolingo, but for X." Maybe X is medical Spanish for nurses, maybe it is sign language, maybe it is coding for kids. The first thing that founder usually reads is an agency page saying the whole thing costs somewhere between $40,000 and $200,000. Five times wide, with nothing underneath it.

So here is the method instead. List the features the real product has, using Duolingo's own filings and blog. Split what a v1 needs from what the company built over more than a decade. Price each feature in hours. Then price the thing everyone forgets, the course, as a separate line. Finally, read the running costs straight off the vendors' own price pages. Every number below is either cited, one of our own published rates, or labelled as my arithmetic.

One thing surprises people every time. The part that looks hard, speech recognition, is cheap and rented. The part that looks easy, writing lessons, is where the calendar goes.

The published agency ranges for a Duolingo clone cannot be reproduced from anything on the pages that print them, so I cite them only as examples of the claim. Here are four I read on September 30, 2026.

EngineerBabu says an MVP can cost $25,000 to $60,000 and a full-featured app $80,000 to $250,000 or more. There is no hour count behind either band and no feature priced in hours, so there is nothing to multiply and nothing to check.

Glancegives £50,000 for a basic version up to well over £500,000, a tenfold spread. To its credit, the page does say the most expensive part is the content rather than the technology, which I agree with. But the same page also says Apple takes 30% of everything you earn and Google does the same. Apple's own Small Business Program charges 15% to developers with up to $1 million in prior year proceeds, and Google lists auto-renewing subscriptions at 15%. When a page gets the one number you can check wrong, I stop trusting the ones you can't.

OnGraph puts a Duolingo-like app at $50,000 to $250,000. Intelegain says $40,000 to $200,000, and an MVP at $15,000 to $20,000. Neither shows hours or a rate.

The same goes for the statistics these pages like to open with: the size of the language learning market in some future year, how fast downloads are growing, what percentage of learners quit. I looked for primary sources behind them and did not find any I could fetch, so none appear here. The only company numbers in this article come from Duolingo's own SEC filing and its own blog.

Look at the $15,000 to $20,000 MVP figure for a second. At $150 an hour, $20,000 buys about 133 hours. That is about three fifths of my low estimate for the lesson player alone. You could build a quiz app for that. You could not build the thing the page is named after.

Duolingo is a company that spent $306.3 million on research and development in 2025, and that number is the right place to reset expectations. Its 10-K for fiscal 2025reports that R&D expense rose $71.0 million, or 30%, from $235.3 million in 2024, on total revenue of about $1.04 billion.

The same filing says the team was more than 900 employees at December 31, 2025, including more than 430 engineers. It reports over 250 language courses, plus courses in math, music and chess. For the fourth quarter of 2025 it lists 133.1 million monthly active users, 52.7 million daily active users and 12.2 million paid subscribers at period end, with about 9% of monthly users paying.

Now let's do the arithmetic out loud. $306,300,000 divided by $225 an hour is about 1.36 million hours in one year. My high estimate for a lean v1 is 2,080 hours. That is about 0.15% of one year of Duolingo's R&D at our top rate. At $150 an hour the R&D line would buy about 2.04 million hours, which makes the slice smaller still.

R&D expense is not all engineering hours. It includes stock compensation, facilities and other costs, and it covers the English Test, math, music and chess as well as languages. The point is scale, not precision. You are not building Duolingo. You are building the smallest learning product someone will open every day and eventually pay for.

Here is the feature inventory that matters, split by what a v1 needs.

Two items on the second list deserve a note. Duolingo described Birdbrain on its blog in October 2020as a model that learns both how proficient each learner is and how hard each piece of material is, and uses that to guess whether a learner will get an exercise right. That is a data science program, not a feature. It only works after you have millions of answered exercises to learn from. A v1 doesn't, so a v1 uses simpler rules.

The AI features are the other. The 10-K names Roleplay, Explain My Answer and Video Call as AI-powered, and notes that they need more spending on computing infrastructure. You can build a version of these on a hosted language model today, and I price that in the fuller launch. But be clear that each conversation has a per-token cost that a flashcard does not.

The gamification layer is what makes Duolingo a daily habit, and it costs more to build than people expect because every rule touches time zones, money and fairness. Duolingo's 10-K says that at the end of 2025 about 43 million daily users had a streak of seven days or longer and about 15 million had a streak of a year or longer. Those numbers tell you how central the mechanic is.

Streaks. A streak counts the days in a row a learner finishes a lesson. Simple, until you ask what "a day" means for someone who flies from Los Angeles to Paris, or practices at 11:58 pm. Then add the Streak Freeze. Duolingo wrote on its blog that letting learners equip up to two freezes at a time increased daily active learners by 0.38% in relative terms. A freeze is a purchasable item, so now the streak touches your in-app currency and your store receipts too. I price streaks with freezes and time zone handling at 90 to 150 hours, bundled with XP and the daily goal.

Leagues. According to Duolingo's own explanation, there are 10 leagues with Diamond at the top, a new league starts every Sunday in each learner's time zone, and learners are grouped with people who have similar study habits and a similar time zone, even if they study different languages. Building that means a weekly job that forms cohorts, a leaderboard that updates as XP comes in, promotion and demotion at the end of the week, and some protection against people farming XP. I price it at 90 to 150 hours.

Hearts, now Energy. For years free learners lost a heart for each mistake. In a post dated July 3, 2025, Duolingo said it was replacing Hearts with Energy, which goes down as you do lessons, comes back when you answer several questions correctly in a row, and fully recharges in about a day. Its stated reason: beginners were twice as likely to run out of hearts mid-lesson. For a clone the lesson is that the mistake limit is a pricing lever, not a teaching feature. It is what pushes free users toward ads or a subscription. I price a simple version at 30 to 50 hours and would expect to change the rules after launch.

There is a design warning too. If your learners include children, the same mechanics that keep an adult practicing can look very different to a regulator or a parent. I come back to that in the COPPA section.

The lean v1 adds up to 1,320 to 2,080 engineering hours in my estimate. That assumes one course, iOS and Android from one cross-platform codebase, a web based authoring tool for your lesson writers, and a senior team. Every row is my scenario estimate, not a quote.

Let me check the sums out loud, because a table that doesn't add up is the same sin as the agency ranges. Low column: 60 + 220 + 80 + 90 + 30 + 90 + 80 + 40 + 140 + 80 + 40 + 70 + 40 + 120 + 140 = 1,320. High column: 100 + 340 + 130 + 150 + 50 + 150 + 130 + 70 + 220 + 130 + 60 + 110 + 60 + 180 + 200 = 2,080.

Now the money. 1,320 hours at $150 is $198,000. 2,080 hours at $225 is $468,000. So the lean v1 is about $198,000 to $468,000 in engineering. For the middle cases, 1,320 hours at $225 is $297,000 and 2,080 hours at $150 is $312,000.

A few notes on where the hours go, because the table hides the interesting part.

The lesson player is the biggest single line.220 to 340 hours for six exercise types: multiple choice, building a sentence from a word bank, matching pairs, listening and typing, filling a blank, and speaking. Each type is its own small interface with its own grading rules. Grading is where the time hides. Is "I am going to the store" right when the answer key says "I'm going to the store"? Does a missing accent count as a typo or a mistake? A lesson engine that marks correct answers wrong loses users faster than any missing feature.

The authoring tool is the second biggest. 140 to 220 hours. Your lesson writers need to create exercises, attach audio, set accepted answers, preview a lesson as a learner, and publish without calling an engineer. Skip this and every lesson change becomes a developer ticket, which is the most expensive way to write a course.

Gamification is three lines, not one. Streaks and XP, the energy limit and leagues come to 210 to 350 hours together. That is more than the speaking feature, which surprises founders who assume the AI part is the hard part.

Speaking is cheap to build because you rent it.80 to 130 hours covers recording audio on the phone, sending it to a speech API, comparing the result with the expected sentence and deciding what counts as close enough. The recognition itself is someone else's model.

Subscriptions are their own project. 80 to 130 hours for the paywall, free trial logic, receipts from both stores, the server notifications that tell you when someone cancels or renews, and the entitlements that decide who sees ads. We cover the details in our guide to in-app purchase implementation.

Mobile is built once for both platforms. The whole lean table assumes one cross-platform codebase, which is how we'd usually approach it in our React Native development work. Two fully native apps would add a large share of the player and gamification hours a second time.

What about time? 1,320 to 2,080 hours with three or four senior people working in parallel is usually five to eight months including design and testing. That's my estimate from the hours, not a promise, and it stretches when the scope does.

The course content is the part of a Duolingo-like budget that agencies leave out, and over the life of the product it is usually the larger bill. An app with a perfect lesson engine and no lessons is an empty bookshelf.

Duolingo's own history makes the point. When it announced 148 new language courses on April 30, 2025, the company said its first 100 courses had taken about 12 years to develop, and that the new ones came from generative AI, a "shared content" approach and internal tooling (Duolingo investor relations). Shared content means building one strong base course and adapting it to many language pairs, rather than writing each from scratch.

Even with AI, Duolingo describes a human process. In a June 2023 post it explained that its teaching experts write a prompt with fixed rules (for example, an exercise under 75 characters, at A2 level on the CEFR scale, using a specific tense), the model drafts exercises, and the experts review and edit what comes out. Before that, it says, every exercise was written, reviewed, edited and translated by human experts.

So here is my content scenario, labelled as mine. One beginner course with 5 units of 10 lessons each, 50 lessons in total. Each lesson takes 4 to 8 hours to plan, draft with AI help, review for accuracy, set the accepted answers, generate audio and test on a phone. 50 x 4 = 200 hours. 50 x 8 = 400 hours.

If you price those hours at our rate for simplicity, 200 hours at $150 is $30,000 and 400 hours at $225 is $90,000. So one course is roughly $30,000 to $90,000 in my scenario. Your content might be written by a subject expert at a different rate, and in that case use their rate. The hours are the part to plan around.

Then multiply. Each new language pair or subject is another course. The shared content approach lowers the cost of the second course, but only if you designed the first one to be reused. That is a decision for week one, not month twelve, and it is one more reason the authoring tool line in the table matters.

Here are the four scenarios side by side, with one course included where marked.

Checking the combined rows: 1,320 + 200 = 1,520 and 2,080 + 400 = 2,480. For the fuller launch, 2,160 + 200 = 2,360 and 3,490 + 400 = 3,890. At $150, 2,360 hours is $354,000. At $225, 3,890 hours is $875,250.

One more point for founders in regulated subjects. If your course teaches medical, legal or safety content, the review step gets longer, because a wrong answer is no longer just a bad lesson. Budget the high end of the range.

A fuller launch adds about 840 to 1,410 engineering hours on top of the lean v1, for a total of 2,160 to 3,490 hours. Again, my scenario estimate.

Checking the added rows: 60 + 120 + 120 + 60 + 80 + 40 + 160 + 40 + 60 + 100 = 840 low, and 100 + 200 + 200 + 100 + 140 + 70 + 260 + 80 + 100 + 160 = 1,410 high. Add the lean v1 and you get 1,320 + 840 = 2,160 and 2,080 + 1,410 = 3,490.

In dollars: 2,160 hours at $150 is $324,000, and 3,490 hours at $225 is $785,250. The fuller launch is about $324,000 to $785,250 before content. The middle cases are 2,160 x $225 = $486,000 and 3,490 x $150 = $523,500.

Three of these rows deserve a warning label.

Adaptive review. 120 to 200 hours gets you a spaced repetition engine that brings back words a learner is about to forget. It does not get you Birdbrain. A personalization model needs data, and at launch you have none. Start with rules, log every answer, and revisit when you have enough history to learn from.

AI explanations and roleplay. 120 to 200 hours to build on a hosted language model, with prompts, guardrails, a way to review bad outputs and a spending cap. The running cost is per conversation, so this is the feature that belongs in a higher paid tier, which is how Duolingo sells its own AI features in Max.

Age assurance. 40 to 80 hours to read the age category signals that the app stores now provide in some states, ask for parental consent again when your app changes significantly, and react when a parent revokes it. This row did not exist in estimates a year ago. It does now, for reasons covered below.

Speech recognition for a language app costs hundreds to low thousands of dollars a month at a realistic early scale, and text to speech for lesson audio costs almost nothing. I read every price below from the vendor's own page on September 30, 2026. Check them again before you sign.

Google Cloud Speech-to-Text lists V2 standard recognition at $0.016 a minute for the first 500,000 minutes a month, then $0.01 up to 1 million, $0.008 up to 2 million and $0.004 above that. The older V1 API gives 60 free minutes a month, then $0.016 with data logging or $0.024 without (Google pricing). Data logging means letting Google use your audio to improve its models, which you should think hard about if your users are children.

Azure AI Speech is the one built for language learning. Its pricing page describes pronunciation assessment as a feature that gives speakers feedback on accuracy and fluency, and says it is charged as standard speech to text. The page loads its prices with a script, so I read the numbers from Microsoft's public Retail Prices API for East US: $1.00 per audio hour for real-time speech to text, $0.18 per hour for batch, and $0.30 per hour for the enhanced add-on features, which the pricing page lists as including prosody scoring. The free tier includes 5 audio hours a month.

AWS Transcribe in US East is $0.0001667 a second for streaming, about a cent a minute, and $0.0001 a second for batch, billed in one-second increments with no minimum for standard transcription (AWS pricing).

Deepgram lists Nova-3 monolingual streaming at a current promotional $0.0048 a minute on pay as you go, with a regular price of $0.0077 (Deepgram pricing).

Here is the catch that the price table can't show. Google, AWS and Deepgram give you back a transcript: what the learner said. For a language app you usually want to know how well they said it. Azure's pronunciation assessment returns scores. With the others you compare the transcript with the expected sentence and accept or reject, which works for "did they say the right words" and fails for "did they roll the r." Pick the API by the answer you need, not by the cheapest minute.

A worked monthly bill.Suppose the app reaches 10,000 daily active learners, and each one does five speaking exercises a day of about five seconds each. That's 10,000 x 25 seconds x 30 days = 7,500,000 seconds, or 125,000 minutes a month. This is a labelled usage scenario, not a forecast.

Walk through the ones that need explaining. Azure bills per hour, so 125,000 minutes divided by 60 is about 2,083.3 hours, which at $1.00 is $2,083.33 and with the $0.30 add-on is $2,708.33. AWS bills per second, so 7,500,000 x $0.0001667 is $1,250.25. I did not model any rounding a vendor applies to very short clips, and five second clips are exactly where rounding rules matter, so read each vendor's billing notes before you trust the smallest number.

Text to speech. Every sentence in a lesson needs audio, and you generate it once, not per learner. Google lists Neural2 voices at $16 per million characters with the first million free each month, and its newest Chirp 3 HD voices at $30 per million, also with a million free (Google TTS pricing). Amazon Polly lists neural voices at $16 per million characters and generative voices at $30 (Polly pricing). Azure's price list shows neural voices at $15 per million.

Do the arithmetic for one course. 50 lessons x 150 sentences x 60 characters = 450,000 characters. On Google Neural2 that sits inside the free million. On Polly neural, 0.45 x $16 = $7.20. On Azure neural, 0.45 x $15 = $6.75. Store the files on Amazon S3 at $0.023 per GB a month and a few hundred megabytes costs pennies. Voicing a whole course costs less than lunch.

A new language app selling subscriptions will usually pay the app stores 15% of in-app revenue, not 30%, as long as it qualifies for the reduced rates. Duolingo's own 10-K says it paid Apple and Google generally 15% to 30% of in-app payments in 2025, and that 62% of its revenue came through the Apple App Store and 20% through Google Play. That concentration is normal for this category.

Apple. The Developer Program costs 99 USD a year, and the standard commission on digital goods is 30%, or 15% in the Small Business Program (Apple). The Small Business Program is open to developers with up to $1 million in proceeds in the prior calendar year and to new developers. Separately, for auto-renewable subscriptions Apple pays 70% in a subscriber's first year and 85% after that subscriber has a year of paid service (Apple subscriptions).

Google. Registration is a one-time $25 (Google Play help). The service fee page lists auto-renewing subscriptions at 15%. It also describes a new structure for users in the US, UK and EEA starting June 30, 2026, where auto-renewing subscriptions are 10% plus a 5% billing fee when the purchase goes through Google Play billing (Google Play service fees). Either way, it comes to 15% on a subscription billed through Google Play.

A worked example.Suppose 1,000 subscribers pay $9.99 a month, all on iOS. That is $9,990 a month in gross subscriptions. In the Small Business Program, Apple keeps 15%, or $1,498.50. Outside it, in a subscriber's first year, Apple keeps 30%, or $2,997. The difference is $1,498.50 a month, or $17,982 a year, for filling in one enrollment form. Enroll before launch.

Free trials, family plans and a higher AI tier all multiply the subscription logic. Duolingo's filing mentions a family plan for up to six subscribers and a Max tier priced above Super. Each is another product in both stores, another set of server notifications and another thing QA has to test on both platforms.

If children can use your learning app, the amended COPPA Rule has been enforceable since April 22, 2026, and it changes what you have to build. This section describes obligations, not legal advice. I am not a lawyer, and a kids' product needs one.

The status. The FTC published the amended Children's Online Privacy Protection Rule in the Federal Register on April 22, 2025. It took effect June 23, 2025, and regulated companies had until April 22, 2026 to comply, apart from a few provisions aimed at safe harbor programs (Federal Register). That date has passed. The rule is in force.

COPPA covers services directed to children under 13 and general audience services with actual knowledge they collect personal information from children. Duolingo's 10-K describes its own approach: the app is aimed at a general audience, and anyone in the US who says they are under 13 at registration gets an ad-free version with special enrollment and restricted information practices. That is a useful model for a clone, and it is why my lean table has a 70 to 110 hour line for an age gate, a restricted under 13 mode and data retention jobs.

Four parts of the amended rule matter most for a language app.

• Voiceprints are now biometric identifiers inside the definition of personal information. An app that listens to children speak has to treat that audio carefully.
• There is a narrow exception for audio: you can collect a file with a child's voice and no other personal information to respond to the child's specific request, if you use it for nothing else, don't disclose it and delete it immediately after responding.
• You need a written information security program and a written data retention policy that says why you keep children's data and when you delete it.
• Disclosing a child's personal information to third parties, such as ad networks, needs separate verifiable parental consent, and you can't make access to the service depend on it.

Put those together and the build consequences are concrete. Speaking exercises for under 13 users should send audio for scoring and throw it away, not store it for progress history or model training. That also argues against the Google data logging option for those users. Ads, if you have them, should be off for children, which is what Duolingo says it does. And the retention policy has to be real code: scheduled jobs that actually delete data on the timetable the document promises.

App store age assurance. A new layer arrived in 2026, and its status has moved fast. Texas SB 2420 requires app stores and developers to handle age assurance and parental consent for minors. A district court enjoined it in December 2025, and Apple paused its plans (Apple, December 23, 2025). On June 3, 2026 Apple said a court ruling had lifted the injunction, and that from June 4, 2026 new Apple Accounts in Texas get age assurance and parent or guardian consent for minors under 18 for downloads, in-app purchases and significant app changes (Apple, June 3, 2026). Developers read the age category through Apple's Declared Age Range API and use a Significant Change API to trigger new consent.

Apple also says it shares age categories for new accounts in Utah from May 6, 2026 and in Louisiana from July 1, 2026 (Apple Developer News). California's AB 1043, the Digital Age Assurance Act, was chaptered on October 13, 2025 and becomes operative January 1, 2027. It requires operating system providers to ask for a user's age at account setup and send an age bracket signal to apps (California Legislative Information).

What does that force you to build? A way to read the age signal on each platform, a decision about what a 13 year old sees versus a 17 year old, a flow for asking parents again when you ship a significant change, and handling for a parent who revokes consent. That is the 40 to 80 hour row in the fuller launch. I found no public dollar figure for compliance with these laws, so I won't invent one. The cost is the hours, plus a lawyer's review that I can't price for you.

If you sell to schools, the rules change again, and student records law enters the picture. We wrote about that side in FERPA issues in vibe coded edtech apps.

Year two costs upkeep, new content and usage, and the content line is the one that decides whether learners stay. The usage side you can read off the tables above: at my 125,000 minute scenario, about $600 to $2,700 a month for speech depending on vendor and whether you buy pronunciation scoring.

The upkeep is steady. Apple and Google ship new OS versions every year. Store billing libraries get new required versions. Age assurance rules keep changing state by state. Each is a small project.

My scenario for engineering upkeep: 40 to 60 hours a month. At $150 an hour, 40 hours is $6,000. At $225, 60 hours is $13,500. So about $6,000 to $13,500 a month, or $72,000 to $162,000 over a year, before new features.

My scenario for new content: 10 new lessons a month at 4 to 8 hours each, so 40 to 80 hours. At $150 that is $6,000 at the low end; 80 hours at $225 is $18,000. Over a year that is $72,000 to $216,000. A learner who finishes your only course has no reason to keep a streak.

Add the two and year two runs about $144,000 to $378,000 in hours, plus usage. That's my estimate, not a benchmark, and it is the number most founders never put in the deck.

If you add the AI tier, add its per-token cost as a usage line and set a hard monthly cap before launch. We walk through the full ownership math, including the years after launch, in app total cost of ownership.

Other products in this series have a very different year two. The Calm cost breakdown is also a content business, but its content is recorded audio, so its running bill is storage and streaming rather than speech recognition.

A real v1 of a Duolingo-like product is bigger than our starter packages, and the arithmetic says so before I have to.

Our MVP development page publishes three bands: $15,000 to $25,000, $30,000 to $50,000, and $55,000 to $75,000 plus. At $150 an hour, $75,000 buys 500 hours. My lean v1 starts at 1,320 hours. So even the top package covers a little over a third of the smallest version of this product.

What does fit inside a package is a narrower first test. For instance: one short course, a lesson player with three exercise types, a streak and XP, a paywall, and no leagues or speaking yet. That is enough to learn whether people come back on day seven, which is the question that decides everything else.

We have built learning products before. For Clinical Edify we built a healthcare education platform with video courses, certification tracking and continuing education credits, using React with TypeScript, Node.js, PostgreSQL, Vimeo for secure video hosting and Stripe for payments. It is not a gamified language app, but the lesson carries over: we rented the media layer and spent our hours on the parts specific to the learner. More background on the category is in our educational app development guide.

The terms that stay constant: a senior build rate of $150 to $225 an hour, full source code and IP transfer to you on full payment, and a 30-day post-launch warranty. We've been building since 2016, first in France and as a US company since 2019, from Los Angeles.

What we won't do is quote you $40,000 for "Duolingo" and then discover the other 1,300 hours, and the whole course, in change orders.

The biggest red flag in a quote for a Duolingo-like app is a total with no hours behind it. Here are the others I'd look for.

• No line for content. If the quote doesn't say who writes the lessons, how many and in how many hours, the course is either missing or hidden.
• No authoring tool. Without one, every lesson edit becomes a paid developer ticket for the life of the product.
• A promise of a Birdbrain-style personalization model at launch, before you have any learner data to train it on.
• Speaking exercises priced with no named speech API, or with a transcript-only API when the pitch promised pronunciation scoring.
• A claim that Apple and Google take 30% of everything, which ignores Apple's Small Business Program and Google's 15% subscription rate.
• Nothing about COPPA, voice recordings or data retention for an app that children might use.
• No mention of Texas, Utah or Louisiana age assurance for an app with teenage users.

One more that is subtler. If the quote prices gamification as a single line called "gamification," ask for it broken into streaks, the mistake limit and leagues. Each has its own rules, its own edge cases and its own tests. A single line usually means nobody has thought about time zones yet.

Several things in this article rest on less than a full primary check, and you should know which ones.

• Every hour figure, including the content hours, is my scenario estimate. None is a measurement of a real project or of what Duolingo spent.
• Duolingo does not publish what a course costs to make. The 12 year figure for its first 100 courses comes from its April 2025 announcement, which I confirmed through search results because the investor site blocks automated reads.
• Azure's pricing page renders prices with a script, so the Azure figures come from Microsoft's Retail Prices API for East US.
• I did not model how each speech vendor rounds very short clips, which can change the bill for five second exercises.
• I relied on Apple's own developer notices for the Texas, Utah and Louisiana timelines rather than the court filings themselves.
• I found no public dollar figure for complying with COPPA or the app store age laws, so none appears here.
• Vendor prices were read on September 30, 2026 and change often.

Three steps, and none of them need a developer.

• Write ten lessons of your course by hand, with accepted answers. Time yourself. Multiply by the number of lessons you need, and you have your real content budget.
• Decide whether children under 13 can use the app. That one answer changes the speech vendor, the ads and the data retention design.
• Estimate your speaking minutes: daily learners x exercises a day x seconds each x 30, divided by 60. Multiply by $0.016 and by $0.0048 to bracket the speech bill.

Then take the feature list and the ten lessons to any developer, us included, and ask for hours per feature and hours per lesson. If they won't give you hours, you've learned something.

Time to get to work.