How do AI apps make money? The revenue models

How do AI apps make money?

AI apps make money mostly from monthly subscriptions, plus business seats, per-token API sales and, since February 2026, ads in ChatGPT's cheaper plans. The largest still spend more than they earn and cover the gap with investors' money. Plainly earns from one line: a $5 subscription.

Every answer an AI app gives costs money to compute, on expensive chips, every time. A heavy user costs the company far more than a light one. Each revenue model below is a way of matching that running cost to someone who pays for it.

Seven sources of money, with real examples

Sources: OpenAI's pricing page (US prices, 3 October 2026), claude.com/pricing (11 October 2026), one.google.com, Search Engine Journal (February 2026).

The last two rows bring in no sales directly. Training data makes the next model better, and investor money pays the bills while revenue catches up with costs.

Subscriptions pay for most consumer AI

A flat monthly plan sits at the center of every big consumer AI app. It works while the average subscriber costs less to serve than they pay. Heavy users are the risk. In January 2025 Sam Altman wrote that OpenAI was losing money on its $200 ChatGPT Pro plan because "people use it much more than we expected" (TechCrunch, 5 January 2025).

Usage limits are how a subscription protects itself from that. I pay $100 a month for Claude Max 5× and still hit its usage limit. Before that I was on Claude Pro for a long time, and my friends and I always ran out. The limit keeps a flat price safe for the company, and it lands hardest on the people who use the product most. How Claude's usage limits work has the mechanics.

Ads arrived in ChatGPT in 2026

OpenAI began showing ads to logged-in adults on ChatGPT Free and Go in the US on 9 February 2026. They sit below the answer, labeled as sponsored. OpenAI says advertisers get aggregate views and clicks, never your conversations, and that ads stay out of health, mental health and political chats (Search Engine Journal, February 2026). At launch, Free users could switch ads off in exchange for fewer daily messages. Go users lose them only by upgrading to Plus.

On 6 October 2026 OpenAI announced visual ads shown beside images that ChatGPT generates, for Free and Go users, with a US test starting later in October (The Star, 8 October 2026). Anthropic went the other way and pledged in February 2026 to keep Claude ad-free. The plan-by-plan detail is in which ChatGPT plans have ads.

API sales and business seats

The labs also sell their models by the token, a chunk of text about three-quarters of a word long. Developers pay per million tokens sent in and per million generated, and the spread between models is wide. On OpenRouter's list on 11 October 2026, Claude Fable 5.1 cost a hundred times as much per output token as GLM-5.3 Flash, an efficient open-weights model from Z.ai.

Business seats are subscriptions bought by employers. Claude Team costs $25 a seat per month, or $20 billed annually, with a $125 premium seat for heavier use. Enterprise adds usage billed at API rates (claude.com/pricing, 11 October 2026). Business plans are also kept out of training by default: Anthropic's 2025 change to consumer data excluded Claude for Work, Education, Government and the API (TechCrunch, 28 August 2025).

Partner bundles

Some AI subscriptions reach people through another company's product. My ChatGPT Go came included with my Revolut subscription. I also had a free year of Perplexity Pro through my telco, Optus. I used it heavily for a few weeks, then not at all for ten months, and I was glad I hadn't paid for it. The terms between the companies aren't public. For the AI app, a bundle adds subscribers who never went looking for it.

Your chats as training data

Consumer conversations are raw material for the next model. ChatGPT uses your chats for training unless you turn it off, on every consumer plan from Free to Pro (OpenAI's pricing page). Anthropic moved consumer Claude, including paid Pro and Max, to the same opt-out setup in 2025, and keeps chats for five years if you leave training on (TechCrunch, 28 August 2025). DeepSeek's own app trains on what you send, stores data in China and takes opt-out requests by email (DeepSeek privacy policy, updated 10 February 2026).

Paying for a consumer plan leaves training switched on until you change the setting yourself. The company gets no cash for your chats. It gets better models, which keep the subscriptions and API sales coming.

Investor money covers the gap

OpenAI took in $13.1 billion in revenue in 2025 and still lost money. On 31 March 2026 it closed a single $122 billion funding round at an $852 billion valuation. A company with that backing can price below cost for years while investors cover the difference. A small app funded the same way has a harder choice when the money runs out: raise the price or cut what the plan includes. Why ChatGPT is so expensive covers what the big labs spend.

Why is there no free plan?

Some AI apps skip the free plan because a free plan always has a payer somewhere else: advertisers, your data, investors or the subscribers on the paid tiers. An app that runs no ads and doesn't train on chats is left with its paying customers.

My own view is that free AI plans don't pay for themselves, so they run ads or take your data, to sell or to train on. For some people that trade is fine. I don't find free apps nice to use. There's an old line about free online services that fits: if you are not paying for the product, you are the product.

ChatGPT's free plan is good, and you can't complain about free. I still don't think you should rely on a free AI product.

Plainly has no free plan. New accounts can ask a limited amount free each month, so you can see it work before you pay. Subscribers pay for that allowance, the same way paid users fund any capped free tier. It is small on purpose: enough to try Plainly, too small to rely on.

How a $5 AI app makes money

Plainly uses the first row of the table and nothing else. The $5 subscription is the only revenue line. There are no ads, chats are never used to train AI, and there is no free plan to fund beyond the small monthly allowance.

The price works because of the model underneath. Plainly runs one capable open-source model and switches when a better one comes out. You never pick. The team chooses a model that answers everyday questions well and costs a small fraction of a frontier model to serve. The OpenRouter prices above show the size of that gap between the two classes. Open-weights models can be run by many hosts competing on price, which keeps them cheap to use.

In my view $5 is sustainable. It won't make Plainly millions, but it is a good service that still makes some money on top.

The trade-off is a narrower product. Plainly writes, plans, explains and gives second opinions. It searches the web when a question needs current information and links its sources, and it reads documents and photos. It has no image or video generation, no projects and no memory across chats. What's the catch with cheap AI goes through the full list, and does Plainly train on your chats covers where a message goes.

Will Plainly raise its price or shut down?

The cost Plainly depends on has been falling fast. Epoch AI found in March 2025 that the price of reaching a given level of AI performance was dropping between 9 and 900 times a year, depending on the task. An app built on efficient open models benefits each time a better one ships at a similar price.

The risk is that trend stopping. If efficient models stopped improving or got dearer, every cheap app would feel it, Plainly included. Apps priced below cost on investor money are the ones most exposed when funding stops, and Plainly's revenue is its subscriptions. Billing is monthly, and you can cancel any time in settings.