Is it worth building an AI companion app in 2026?

How real is the demand for AI companions?

Realer than anything else in consumer AI. maybe worth building pulled the receipts: companion apps had racked up 220 million lifetime downloads by July 2025, with downloads up 88% year over year, per Appfigures data reported by TechCrunch in August 2025. Character.AI users average 75 minutes a day on the app, per Sacra. That's not a tool. That's a relationship.

If you want to know how deep it runs, look at what happened when OpenAI retired GPT-4o with no warning in August 2025. MIT Technology Review interviewed users who had built relationships with the model. One, who goes by Starling, put it this way: "I've grieved people in my life, and this, I can tell you, didn't feel any less painful. The ache is real to me." Another wrote on Reddit that the replacement model was "wearing the skin of my dead friend."

So demand isn't the question. People bond with these products harder than with almost any software ever shipped. The question is why almost none of that turns into money.

If the demand is this real, why is the money so thin?

Start with the headline number, because it's small. Appfigures tracked $82 million in consumer spending on companion apps in the first half of 2025, on pace to clear $120 million for the year. For the most talked-about category in consumer AI, that's a small pot. And the split is brutal: 337 apps generate revenue, 128 of them launched in 2025 alone, and the top 10% take 89% of everything. The other 300 or so apps divide about $13 million, which averages out near $44,000 each before Apple and Google take their cut.

The leader makes the problem plainer. Sacra pegged Character.AI at $30 million annualized in July 2025, projected to hit about $50 million by year end, on roughly 20 million monthly users. Run the math and you get about $2.50 per user per year. The subscription costs $9.99 a month, so almost nobody pays. Seventy-five minutes a day of attention, $2.50 a year to show for it.

Why the gap? The core audience skews young and pays like it. The thing users would pay most for, unfiltered intimacy, is exactly what app stores and mainstream payment processors punish; Stripe and PayPal won't touch NSFW, so the highest-paying slice of the market runs through high-risk adult processors with worse rates. The category's revenue per download did double from $0.52 in 2024 to $1.18 in 2025, so monetization is improving. From a very low floor.

Who already owns the space?

Character.AI owns generic persona chat, and even it needed a lifeline: Google paid about $2.7 billion in 2024 to license its tech and hire back its founders. Replika has been running since 2017 and survived its own users revolting over feature changes. And the labs now ship companionship natively: xAI launched animated companions inside the Grok app in July 2025, behind its $30-a-month SuperGrok tier, and 68% of Grok users tried the mode in week one.

That last one is the tell. When the model owners treat companionship as a feature of the model app itself, a standalone companion running on a rented model is the same trap we flagged in the ChatGPT wrapper verdict: your product is one platform decision away from being a default setting. Except here it's worse, because your users are emotionally attached to a personality you don't control. When the lab deprecates the model, your product's soul changes overnight, and the GPT-4o grief threads show exactly what that does to your user base.

What changed on January 1, 2026?

Regulation showed up. California's SB 243, signed October 13, 2025 and in effect since January 1, 2026, is the first US law aimed squarely at companion bots: mandatory AI disclosure, crisis-intervention protocols when users show signs of self-harm, a reminder to minors every three hours, and a private right of action at $1,000 per violation. Character.AI saw it coming and shut off open-ended chat for under-18s in November 2025, before the law forced it.

Read that against the revenue math. Compliance is now a fixed cost per user, in a market that earns $2.50 per user per year. Big players can amortize lawyers and age verification. A two-person indie companion app cannot, and "we'll worry about it later" is a bad plan in the one consumer AI category where the harm stories make national news.

When is an AI companion app actually worth building?

A few shapes still hold up:

• A companion with a utility spine. The companion is the interface; the job is the product. Language practice, eldercare check-ins, structured coaching. These get bought for the outcome, so they charge like tools instead of monetizing loneliness, and they sidestep most of the $2.50 problem.
• You own the memory, not the model. The GPT-4o grief proves the attachment lives in the persona and the shared history, not the weights. If years of relationship memory live with you and survive a model swap, you own a switching cost no lab can ship in a weekend.
• An audience the big apps won't serve. A specific language, culture, or age-verified adult lane where you price the high-risk payment rails in from day one, instead of discovering them at your first frozen Stripe account.

What's not worth building is the general-purpose persona chat app on a rented model. That market is a lottery where the top tenth of apps already holds 89% of the prize pool and the labs print their own tickets. California now audits the drawing.

The one way we're wrong: this could be early innings, not a ceiling. a16z has ranked companionship among the top consumer AI use cases since 2023, and if per-user revenue ever catches up to 75-minutes-a-day engagement, the winners get enormous. But that upside accrues to whoever already has the users. In 2026, that's not the new entrant.

The test before you build is the same one we apply to every wrapper: find the space receipt (a real company making real money in your specific wedge) and the pain receipt (one real person, in their own words, on why the existing apps fail them). Then answer one question with a number: who pays you, and why doesn't the $2.50 problem apply to them? The broader read on crowded consumer AI is in our saturation verdict: the commodity layer is full, the workflow layer isn't.

The market's own data says it best: 337 apps competed for $120 million in 2025, and the top tenth of them took 89% of it.

Frequently asked questions

How much money do AI companion apps actually make?

Appfigures tracked $82 million in the first half of 2025, on pace to clear $120 million for the year. 337 apps generate revenue and the top 10% take 89% of it. The remaining 300 or so split about $13 million.

How much revenue does Character.AI make per user?

About $2.50 per user per year: roughly $50 million in projected 2025 revenue across about 20 million monthly users, per Sacra. That's with users averaging 75 minutes a day and a $9.99 monthly subscription on offer. Engagement is enormous, paying is rare.

Is the AI companion market saturated in 2026?

The generic persona-chat layer is: 128 new companion apps launched in 2025 alone, and 17% of companion apps have "girlfriend" in the name. Narrow companions with a real job attached, for a specific audience, are far less crowded.

What laws regulate AI companion apps in 2026?

California's SB 243, in effect since January 1, 2026: mandatory AI disclosure, crisis-intervention protocols, a three-hour reminder for minors, and a $1,000-per-violation private right of action. Character.AI cut off open-ended chat for under-18s in November 2025, before the law forced it.

Won't the AI labs kill my companion app?

They're already in the market. xAI shipped animated companions inside Grok in July 2025 behind its $30-a-month tier. A generic persona on a rented model is a platform decision away from irrelevance. What a lab won't ship is your niche audience, your distribution, and years of relationship memory a user can't rebuild elsewhere.

What should I check before building one?

Two receipts: a real company earning in your specific wedge, and one real person describing in their own words why current apps fail them. Then name who pays you and why the market's $2.50-per-user-per-year problem doesn't apply.