In 2025, people spent more money in regular apps than in games for the first time. Generative AI is the main reason. But if you're thinking of building "an AI app", the money is not where it looks like it is. Chat assistants are taken. Photo, video, voice, and AI tucked inside a daily habit are where people still pay. I build iOS apps, so I pulled together the numbers by niche and wrote down where a small team can still get in.
Short version: don't build another assistant. Build something that produces a file people keep, or remove the boring step from a task people already do every day. The second one is the realistic bet for a solo developer.
This is part 1 of a short series on how people make money with AI right now. This one is about mobile apps; part 2 covers websites and web services.
The market: there is money
According to Sensor Tower's State of Mobile 2026, consumers spent $167B in apps in 2025, up 10.6%. Non-game app spending grew 21% and passed games, which grew 1.3%. Generative AI apps were downloaded 3.8B times (double the year before), and their in-app revenue nearly tripled to over $5B.
Growth hasn't stopped. Sensor Tower projects AI apps will pass $4B in H1 2026, 36% more than H2 2025. Time spent doubled year over year, from 17.2B to 36B hours. More than 200,000 apps now mention AI in their descriptions.
That last number matters most. With 200,000 apps saying "AI", the label on its own sells nothing. What sells is the job it does.
Chat assistants: the field is taken
In Q1 2026, ChatGPT, DeepSeek and Google Gemini held almost 90% of time spent in AI assistant apps (Sensor Tower). Appfigures estimates ChatGPT made $187M net on mobile in October 2025, with Grok at $8.7M, Claude at $5M and Perplexity at $3.1M. Even companies with their own frontier models are 20–40x behind.
It wasn't always like this. In 2023–2024, thin wrappers did well: Appfigures puts ChatOn at $135M and Chat & Ask AI at $109M over two years. That was before OpenAI and Google shipped good apps of their own, when people searched the App Store for "ChatGPT" and installed whatever came up first. Today the free original beats any wrapper. That's my read, not a finding from the reports.
The niche map
Figures come from different sources and aren't directly comparable: some are net of store fees, some gross, some are company claims.
| Niche | Who earns | Numbers | Barrier for a small team |
|---|---|---|---|
| General-purpose assistants | ChatGPT, Claude, Grok, Perplexity | ChatGPT: $187M net in Oct 2025 (Appfigures) | Closed |
| Video generation & avatars | Higgsfield, HeyGen, Runway, Kling, Synthesia | Higgsfield claims $1B annualized revenue (company claim) | High: generation is expensive, margins thin |
| Photo: editing, retouching, product shots | Photoroom, Evoto, FaceApp, Reface, Topaz | About ten photo tools in a16z's top 50 by spend | Medium: a narrow job works |
| Audio: music, voice, read-aloud | Suno, ElevenLabs, Speechify, Moises | Suno is #7 in a16z's top 50 | Medium |
| Productivity: meetings, notes, slides | Otter, Plaud, Fireflies, Gamma, Notion | 15 of 50 spots in a16z's top 50 | Medium: higher prices, companies pay |
| Companions & roleplay | Zeta, Tipsy Chat, Character.AI | $165M general + $163M romance in H1 2026, gross (Appfigures via Decrypt) | Low technically, high on moderation and reputation |
| Homework & study | Gauth (ByteDance), Question.AI | ~$2.5M net combined in 2024 on 15M installs (Appfigures) | High: big players plus free ChatGPT |
| AI inside a daily habit: food, health, fitness | Cal AI and narrow trackers | Cal AI: $40M+ trailing-12-month sales, acquired by MyFitnessPal in March 2026 | Low–medium: best shot for a solo dev |
The a16z ranking is the top 50 consumer AI products by US consumer spend (October 2026, YipitData), web and mobile combined. More than half of the list is photo, video and audio.
Why people pay for photo, video and voice
My interpretation, but it fits the numbers.
- The result is a file. You read a chatbot answer and close it. A video, an edited photo or a voiceover gets saved, posted, sent to a client. It's easier to pay for something you can take away.
- Free ChatGPT is worse at it. A general assistant can draw a picture. It won't batch-remove backgrounds from forty product shots at marketplace size.
- Professionals pay. Marketplace sellers, photographers, creators and marketers treat these as work tools that pay for themselves.
The flip side is cost. Every video generation costs money, and an all-you-can-eat plan will put you in the red. Credits instead of unlimited, and expensive models only on the paid tier.
Where a solo developer fits: AI inside a habit
The most telling story of the year is Cal AI. Two high schoolers built a calorie counter by photo: snap your plate, get calories and protein. In under two years it passed 15M downloads, did over $40M in sales in the trailing 12 months, and MyFitnessPal bought it in March 2026 (terms undisclosed).
The model inside isn't theirs, and food recognition isn't unique. What worked: AI removed the most tedious step of a task people already wanted to do. Counting calories was always the goal; typing in every spoonful was the problem.
The recipe:
- Pick a task people do regularly, daily or weekly.
- Find the most tedious step: data entry, transcription, counting.
- Let AI remove that step, and keep the rest a good ordinary app.
I do this in my own fasting app: calories from a photo are part of the daily log, and recognition runs on the phone through Apple Foundation Models, offline and with no API bill. When a request costs nothing, a free tier and a long trial don't turn into losses.
A habit also fights the other AI-app problem, churn. People leave AI apps faster than regular ones, and what keeps them is the recurring task, not the model. I covered that in an earlier breakdown of RevenueCat's data.
What can break it
- The platform eats your feature. If your product is one button, it'll show up in ChatGPT, Google Photos or iOS itself. Data, history and a good workflow are what protect you.
- Store rules. In April 2026 Apple briefly pulled Cal AI: it took payments via Stripe outside in-app purchase and showed the weekly price more prominently than the actual charge. It came back after fixes.
- Cost. A cloud model is a variable cost per user. Do the math before launch, not after the first invoice.
- Companions are their own thing. Hundreds of millions in revenue, but the top 10% of apps take 89% of it (Appfigures, 2025), and moderation and reputational risk are the highest anywhere.
What I'd pick
Not an assistant, and not a video generator: the first is taken, the second needs model budgets a solo developer doesn't have. I'd take a boring recurring task in a niche where people already pay for apps (food, fitness, health, bookkeeping, photos for sellers), have AI remove the most tedious step, on-device where possible, and sell an annual plan with a long trial.
Less exciting than "my own ChatGPT". But it's what the one real success story of the year looks like.
Originally published at klukyanov.ru.
Shorter weekly write-ups (in Russian) — on Telegram.
Top comments (0)