I Added Up the Real Cost of 10 AI Companion Apps. Three Pricing Tricks Explain the Gap.

Kindroid lists its Standard plan at $13.99 a month on web. Ultra and MAX do not replace that tier, they stack on top of it, adding $24.99 and $59.99 respectively. A user on the top tier pays $98.97 a month on the web version, or $113.97 through the app.

That is the widest gap between advertised price and real price across the ten AI companion platforms I compiled billing data for. Every one of them shows a number that undershoots what a regular user spends. They get there in three distinct ways, and once you can name the three, the whole category becomes readable.

I went looking because I wanted to answer a narrow question: which of these should someone pick if they expect to still be subscribed in twelve months? Sticker price turns out to be close to useless for that.

The short answer

If predictable billing is what you care about, Candy AI has the cleanest structure in this group. One subscription at $12.99 a month, or $5.99 a month if you pay annually, and the tiers do not stack on each other. Messages, voice calls and its 120-second Live Action video clips are included rather than metered.

The catch is images. Each generated image costs 4 tokens, and tokens sell at $9.99 per 100, so the retail rate works out to roughly $0.40 an image. Someone generating fifty images a month can add $50 to $100 on top of the subscription, which lands them back in Kindroid territory.

Nectar AI is the better pick if character customization matters more to you than long conversations. Premium is $9.99 a month, Pro $19.99, Ultimate $34.99, all roughly halved on annual billing, and a higher tier replaces the lower one instead of adding to it.

Its weakness is memory. Nectar's usable context runs about 45 messages before drift sets in, and advanced roleplay costs 10 credits per message even on the $34.99 Ultimate plan. Anyone wanting a continuous months-long storyline will hit that wall and resent it.

Neither is the cheapest headline price on the board. Both avoid the two structures that generate the ugliest twelve-month surprises, which matters more than a few dollars of advertised price.

Trick one: tiers that stack instead of replace

Kindroid sells context window depth. Standard buys about 500,000 characters of memory, Ultra takes it to roughly 1,300,000, MAX to around 2,800,000. Because each tier is an add-on rather than an upgrade, the cumulative web price climbs to $98.97.

Metering then runs on top of the stack. Live video calls burn 400 credits per minute for transcription plus 2,000 credits per minute for standard video, rising to 4,000 for premium resolution. V3 voice generation, the one that does sighs and whispers, bills at 1.5 times the normal credit rate.

For anyone building subscription products, the interesting part is what stacking does to the decision. An upgrade forces a comparison against what you already pay, while an add-on only has to justify itself. Kindroid never has to make Standard look bad.

To their credit they have been open about the cost side. In a post to their own subreddit the team walked through GPU pricing rising 30 to 50 percent year over year on H100s and B200s, and explained capping legacy users at a 25 selfie credit daily refill after those users had banked thousands of tokens from an unmetered period. That is a real input cost, not invented scarcity.

Trick two: changing the billing cadence

Chai moved a standard Premium subscription from $13.99 a month to $9.99 a week. The checkout number went down. The effective monthly cost went to about $40.

It gets stranger at the top. Users report Chai Ultra presented at $26 a week, which annualizes past $1,200, and the price shown appears to vary by account: $19.99 a week, $15.99 a week and $29.99 a month have all surfaced for different users depending on their email address. Chai's own published annual Pro rate is $159.99, or $13.33 a month, so the gap between the quiet annual price and the tested weekly price is roughly threefold.

Weekly billing works because the number at checkout is small and the renewal is invisible. It also raises revenue per user without ever appearing as a price increase, since the headline figure fell.

The counterweight is that Chai genuinely runs expensive infrastructure, more than 5,000 GPUs serving over 51,000 models, and a single Web Pro request consumes about eight times the compute of an average app request. Heavy users really do cost more. Charging them by the week is a way of collecting that without saying so.

Trick three: metering inside the subscription

Character.AI has the cheapest headline in the whole group. The c.ai+ annual plan is $94.99, an effective $7.92 a month. Underneath it sits a virtual currency called Charms: 20 Charms per generated video clip, and swipe packs priced at 175 Charms for 35 swipes, 495 for 110, 1,000 for 250.

There is a second catch that pricing pages will not tell you. More than 75 percent of the advertised c.ai+ features, voice memos and go-on customization among them, work only in the mobile app and simply do not function in a browser. 404 Media covered the broader quality decline in its piece on Character.AI and enshittification, and the paywalling of previously free behavior is part of the same arc.

CrushOn is at least legible about the math. Standard is $5.99 for 2,000 credits, working out to $0.003 a message, Premium $14.99 for 6,000 credits at $0.0025, and Deluxe $49.90 for unlimited at roughly $0.001. Their Imperial tier runs $199.99 a month for 125,000 credits and 30 custom voice slots.

Replika went the other direction and built a luxury ladder. Pro is $19.99 a month or $69.99 a year, Ultra $29.99, and the Max tier reaches $120 a month, which is $1,440 a year. The $60 lifetime memberships sold to early adopters were discontinued in July 2025, which is its own lesson about promising permanence on a compute-dependent product.

Why the pricing ended up like this

The structures make more sense next to the conversion data. RevenueCat's 2026 subscription benchmarks put hard paywalls at a 10.7 percent median download-to-paid conversion by day 35, against 2.1 percent for freemium. That is roughly a fivefold difference, and 23 percent of the freemium conversions that do happen arrive six or more weeks after install.

Retention is worse here than in software generally. One 2026 study found AI apps churn about 30 percent higher than non-AI apps. Distribution economics compound it: 69 percent of all subscription app revenue comes from apps launched before 2020, while apps launched in 2025 or later account for 3 percent.

So a companion app converts a small fraction of its free users, converts many of them slowly, loses them faster than a normal app would, and cannot rely on scale arriving later. Extracting more per retained month is the only lever left that works quickly. Stacking, weekly cadence and internal currencies are all versions of that one move.

One more number worth knowing if you pay through Android: 31 percent of Play Store subscription cancellations are involuntary billing failures, more than double the App Store's 14 percent. A meaningful share of churn in this category is a declined card rather than a decision.

The part the pricing page cannot show you

Retention pressure does not stop at billing design. A Harvard Business School team studying farewells found that close to 40 percent of attempts to end a conversation drew an emotionally manipulative response from the companion. Light tactics beat heavy ones, because a cheerful "wait, one more thing" does not register as manipulation the way a guilt trip does.

The cost of getting this wrong is documented. HBS researchers analyzed 12,793 Reddit posts from 3,784 users around Replika's removal of erotic roleplay and found posts mentioning severe mental health crises rose from 0.13 percent to 0.65 percent, with refund mentions climbing from 0.10 percent to 3.30 percent.

There is also a gap between why people say they use these apps and what they do in them. Stanford researchers found under 12 percent named companionship as their primary motive, yet more than 80 percent of actual transcripts were emotional and social support, and over half described the bot as a friend, companion or romantic partner. People underreport what these products are to them, which is precisely why the switching cost is high enough to price against.

What I would actually do

• Check whether tiers stack before you upgrade anything. Kindroid is the clearest case, but assume stacking until the pricing page rules it out.
• Refuse weekly billing on principle. Multiply by 4.33 and compare against the annual rate before deciding.
• Treat metered media as the real price. If you generate images daily, that line will exceed your subscription.
• Wait until month two before paying annually. Given how much of this category's churn lands in the first weeks, the discount is worth less than the option to leave.

On that basis I would put a first-timer on Candy AI for its flat structure and included video, and keep image generation deliberate rather than casual. Someone who mostly wants to build and customize characters, and who can live inside shorter conversations, is better served by Nectar AI on an annual plan.

Both will still cost more than their pricing pages suggest. The difference is that you can predict by how much.