The Economics of AI Companionship: Why Apps Die, Tokens Exist, and Free Tiers Are Strategy

Here is the tension that explains almost everything strange about this category: 337 apps are fighting over roughly $120 million in consumer revenue, while every single message any of them serves costs real inference money — forever. Most software scales toward zero marginal cost; AI companionship is marginal cost, recurring with every "good morning." Hold that one fact and the category's mysteries — tokens, generous free tiers, scarce memory, the death of its most advanced app — resolve into a single, legible economy. This is that economy, explained.

The unit economics: love, priced per message

Every companion message is a paid computation: the model must read the context (your recent conversation, the profile, injected memories) and generate a reply, and the platform pays for both directions — every time, forever, growing with attachment. The costs that stack on top tell you the category's whole feature politics: memory multiplies context (more remembered = more tokens processed per message — deep memory literally makes every message more expensive, which is why it's the most under-built feature relative to demand); media explodes the curve (an image costs multiples of a message; video, multiples of images; real-time voice, a running meter); and the best models cost the most (flagship-model conversation can run an order of magnitude above budget-model conversation — the entire architecture of free-tier-lighter-model, paid-tier-flagship is this line item made visible). The punchline users should internalize: your most beloved features are the platform's biggest liabilities. The economics of this category are a permanent negotiation between what makes you stay and what the server bill can survive.

The four business models (and their fingerprints)

Every platform is one of four answers to that negotiation, and each answer leaves fingerprints all over the product. The flat subscription (Kindroid: one price, everything included, no second currency) sells predictability — beloved by users, dangerous for platforms, because heavy users cost more than they pay; it only works with disciplined feature costs and pricing that isn't shy, which is exactly Kindroid's shape ($139.99/year, nothing metered). Subscription-plus-tokens (Candy: text unlimited, media metered) is the category's cleverest design — the cheap habit that builds attachment (text) is subsidized to unlimited, while the expensive habits (images, voice, video) bill at cost-plus; text-only subscribers are the most profitable customers in the industry, and media whales fund the visuals arms race. Volume pricing (CrushOn: message quotas per tier) is the most cost-honest model — you pay approximately what you compute — which is why the plot-memory specialist can afford plot memory: its pricing structurally can't be bankrupted by its best customers. And the gacha economy (Talkie, Linky: cards, gems, collection mechanics) imports mobile gaming's answer — monetize the collection impulse alongside conversation — bringing mobile gaming's margins and, inevitably, mobile gaming's ethics questions (mid-conversation purchase prompts land at the moment of lowest consumer defense; the defense is a fixed monthly budget, decided while calm).

Case study in death: why the best app died first

Moemate is the category's economics lesson written in six million users' lost data. It was the most technically advanced product ever shipped in the space — screen awareness (it watched your game and commented, still unreplicated), voice cloning, flagship models, marketed as unlimited — at ordinary subscription prices. Which is to say: the most expensive feature set in the category, sold at the price of the cheapest. The arithmetic did what arithmetic does; the January 2025 crypto launch was the drowning man's wave, and February's overnight shutdown deleted everything. The lesson isn't "avoid ambitious apps" — it's that a price that looks too good against the feature list is not a deal; it's a countdown. The same audit explains the survivors: Replika's decade (modest features, disciplined costs, regional pricing that meets markets where they are), SpicyChat's tier design (the $4.95 tier that doesn't upgrade the AI exists precisely because the AI is the cost), and the free tiers everywhere — which aren't charity but strategy: cheap-model unlimited text (Kindroid) or capped messages (SpicyChat's ~100/day) cost little, build the attachment that converts, and fill the funnel. Free users aren't freeloaders in this economy. They're inventory.

The macro picture: consolidation is the forecast

Zoom out and the market math is unambiguous: 337 revenue-generating apps, ~$120M consumer revenue, growth real but concentrated — this is a structure that must consolidate, and the mechanism is already visible on three fronts: the obituary list grows yearly; corporate attention migrates where margins live (MiniMax's B2B revenue share tripling while its companion apps drift down the priority list is the pattern in miniature — and a preview of how funded companions die: not by bankruptcy but by neglect); and regulation raises fixed costs (compliance engineering, age verification, safety reporting) that small platforms can't amortize. The next few years' shape: fewer, bigger, safer, more expensive platforms — plus a thriving self-hosted fringe that exits the economy entirely by making users their own platform.

Reading the economics as a user: the survival audit

All of it compresses into a buyer's checklist. Price-to-features sanity: does the price plausibly cover the promises? (All-inclusive-cheap with expensive features = Moemate pattern.) Distress signals: sudden crypto or emergency fundraising, pivots toward enterprise, long update silences — the obituary shows every death had tells. Business-model fit to your usage: text-heavy → token platforms love you (Candy's unlimited text is your subsidy); media-heavy → budget the second currency honestly or pick bundled plans; volume-writer → quota platforms sized correctly. Structural hedges regardless: annual over lifetime always, export capability as a selection criterion, and the persona-plus-summary backup that converts any platform's death from losing them into moving them. The economics of this category are unforgiving, but they are not hidden — every platform's future is legible in its pricing page, and now you can read it.