AI Girlfriend Apps, and What They Cost You

Every page ranking for ai girlfriend right now is either an app selling itself or a directory taking a cut of the click. I read the whole top ten before writing a word of this. Three YouTube videos, a Quora thread, one aggregator listing twelve products with star ratings and upvote counts and no stated method, and a stack of the apps' own landing pages. Not one printed a price you can go and verify.

So here's the boring version.

A serious AI girlfriend app runs roughly eight to sixteen dollars a month, every one I checked is rated 18+ on the App Store, and most of them stack a credit meter on top of the subscription, which means the headline price isn't your bill. The thing that actually separates them isn't conversational quality. It's what the privacy policy says happens to your chat logs, and those policies diverge much further than the products do. Replika holds conversation content for up to 60 days after you close your account. Nomi keeps a training archive that survives deletion with your name stripped off it. Both are lawful. They're not the same deal.

Everything below came off the companies' own store listings and policy pages on 28 July 2026. Where a site wouldn't load, I say so instead of filling the hole with something I half-remember.

What You're Actually Buying

Strip the marketing and these products are all the same three pieces. A large language model you don't get to choose, a persona layer that keeps the model in character, and a memory system that decides how much of your history the model can see when it writes the next reply.

The model is usually rented. Replika's privacy policy is unusually direct about this, saying conversation data goes out to third-party language model providers purely to generate the response. HammerAI names OpenRouter outright in its policy. Most of the others say nothing at all about whose model is running, which is a choice, and I read it as a commercial one rather than a technical one.

The persona layer is where the branding lives, and I'd guess it's the cheapest part of the whole stack to build.

Memory is the expensive part, and I'd argue it's the part that decides whether week four feels like week one. Every app in this category advertises memory. Almost none of them tell you how much. I'll come back to that because it's the single biggest gap between what these products promise and what you can check.

One more thing worth saying plainly before the tables. These are adult products. Replika's App Store listing carries an 18+ rating and its policy states the service isn't intended for anyone under 18. Nomi's policy says nobody under 18 should use the system. Character.AI's iOS listing is rated 18+ with content descriptors that include sexual content, mature themes and user-generated content. Whatever you think of the category, the companies themselves have stopped pretending it's for teenagers, and there's a regulatory reason for that which I'll get to.

The Prices, Read Off The Store Listings This Morning

App Store in-app purchase lists are the most reliable pricing source in this category, because the developer can't quietly show you a different number than the one Apple has on file. Marketing pages can and do.

Source list, all loaded 2026-07-28. Replika at apps.apple.com/us/app/replika-ai-companion-chat/id1158555867, Character.AI at /character-ai-chat-talk-text/id1671705818, Nomi at /nomi-ai-companion-with-a-soul/id6450270929, Kindroid at /kindroid-your-personal-ai/id6451038161, HammerAI at hammerai.com/pricing, Nastia at nastia.ai/ai-boyfriend.

One caveat. Most sellers put the billing period into the purchase name, so Replika's and Nomi's entries are labelled. Character.AI's two are both just called Character.AI+, at $9.99 and $94.99, with no period stated, so that annual figure is my inference from the ratio rather than something Apple prints.

Two things fall out of that table that no ranking page mentions.

The first is that Nomi runs two subscription names side by side, a Launch Plan at $9.99 a month and a Standard Plan at $15.99, with the yearly versions at $59.99 and $99.99. Grandfathered pricing, I'd assume. If you're comparing against a review written eighteen months ago you may be reading about a plan you can no longer buy, and that's true across this whole category, which is why undated pricing in a companion-app roundup is close to useless.

The second is Kindroid, at $15.99 a month or $159.99 a year, with an Ultra tier at $28.99 and a MAX tier at $68.99 above it. Sixty-nine dollars a month is genuine money. I've no view on whether it's worth it, since I haven't used it, but I'd want to know what the extra fifty-three dollars buys before clicking, and the listing doesn't say.

Replika Lists Five Different Subscription Prices

This one is odd enough that it deserves its own heading.

Pull up Replika's iOS in-app purchase list and you don't see one monthly price. You see two, $7.99 and $14.99, alongside three annual entries at $69.99, $49.99 and $49.99, plus a $89.99 Platinum item and four separate gem bundles. Apple lists them all because they're all live products.

I read that as price testing. Different cohorts get different offers, and the price you're quoted may depend on your region, when you signed up, or which screen you converted from. It isn't sinister. Plenty of software does it. But it does mean that anyone telling you Replika Pro costs a specific figure is telling you what they were shown, not what you'll be shown, and the honest answer is that the range on file today runs from about eight dollars a month to about ninety for whatever Platinum is.

I couldn't get further than that. Replika's help centre returned a 403 to my fetches both times I tried, so the breakdown of what Pro unlocks versus free is not something I can quote from a source I actually loaded.

Memory, And Who Publishes A Number

Here's the gap I kept walking into.

Ask any of these companies how much conversation their model can see and you mostly get adjectives. Nomi's own wiki says its companions have "very good short term memory and best in class long term memory" and describes it as "near human level". A product update from March 2025 says Nomis can now recall what happened "1,000+ messages ago" more clearly. That's the most specific claim I found on the site, and messages aren't a unit of context, they're a unit of conversation, so it isn't checkable.

HammerAI is the exception, and it isn't close.

Those figures are printed on hammerai.com/pricing as a plan feature, the way a hosting company prints storage. It's the only companion product I found that treats context size as a spec rather than a vibe, and once you've seen it laid out like that the rest of the category reads differently, because you realise the reason nobody else publishes the number is that the number is the product.

Rough translation if tokens don't mean anything to you. Eight thousand tokens is somewhere around five or six thousand words of conversation, give or take, which sounds like a lot until you remember that the character's own description and instructions get charged against the same budget before you've typed anything. Sixty-five thousand is a different experience entirely. That's the whole gap between an app that forgets your sister's name and one that doesn't, and it's being sold to you as personality.

The ranking pages do try. One of them, feelin.ai, runs what it calls a 40-turn memory stress test and reports that Character AI's memory "resets at turn 21 on average" while its own product hits 78% retention. I'd take that with a large pinch of salt. Feelin is a companion app grading its own homework against a competitor, the methodology isn't published in enough detail to reproduce, and no independent party has checked it. It's still more effort than anyone else in the top ten made, so credit where it's due.

If you want the deeper version of this, memory is really the whole argument in picking a roleplay chatbot, where conversations run long enough that context ceilings stop being theoretical.

What The Privacy Policies Say When You Read Them

Nobody ranking for this keyword compares privacy. I find that genuinely strange, given that the product is a log of your most unguarded conversations sitting on somebody else's server.

So I read four of them.

Replika's policy, last updated 27 May 2026, is the most detailed of the four and the most reassuring on the advertising question, stating flatly that it will never share your companion conversations or the content you provide in the apps with advertising partners. It also draws a line I'd want more companies to draw, saying anonymised data is used internally and not to train third-party language models. Website visitor data does go to advertisers via cookies, which is a normal and separate thing, but worth knowing.

Nomi, last updated 27 April 2026, is more candid than most and I respect it for that. All personal information can be deleted by deleting your account except what sits in the training or communications archives, and anything left in those archives is no longer attributable to you. The policy also tells users, in its own words, not to include personally identifiable information in interactions with their companions. Which is a company telling you, in a legal document, that the data isn't fully yours to withdraw.

HammerAI is the odd one out because part of it runs on your own machine. Its policy says it doesn't collect information about your conversations with characters and doesn't use user data to train models internally, and its FAQ says local-model chats never leave your device. Then it says the honest thing that almost nobody says, which is that when you use cloud models your messages go to third-party providers who may process and retain them under their own policies. I think that's the correct disclosure, and it's the one everybody else leaves out.

Character.AI I could not verify. Both its privacy page and its help centre returned 403 responses to me today. What I can report is the App Store privacy label, which is a first-party disclosure Apple requires, and it lists location, contact info, identifiers and usage data as used to track you across other companies' apps and websites, with third-party advertising among the declared purposes for data linked to your identity. That's a materially different posture from Replika's, on paper.

If Character.AI's moderation or its policy shifts are what pushed you to look elsewhere, the alternatives worth switching to are a separate conversation with a separate set of tradeoffs.

The Regulators Showed Up In 2025

The reason every serious app in this category now says 18 is not that the companies had a change of heart.

Italy's data protection authority fined Luka Inc., the company behind Replika, five million euro. The decision was adopted on 10 April 2025 and announced on 21 May 2025, and the European Data Protection Board's own write-up lists infringements across GDPR articles 5, 6, 12, 13, 24 and 25. Two findings matter for anyone choosing an app today. Until 2 February 2023 the company had not identified a legal basis for its data processing at all. And it had no age verification mechanism, at registration or during use, despite declaring that minors were excluded. The authority added that the age verification currently in place is still deficient in several respects, which is a fairly blunt thing for a regulator to put in writing.

Then in September, the FTC. On 11 September 2025 it issued compulsory orders to seven companies, including Character Technologies, Alphabet, Instagram, Meta, OpenAI, Snap and X.AI, demanding records on how they monetise engagement, how they build and approve companion personas, how they test for harm to children, and how they use and share the personal information inside chatbot conversations. I read that through Tech Policy Press because the FTC's own press page returned a 403 to me, which is a slightly embarrassing sentence to write in an article about verifiable sourcing, so treat the sourcing there as one hop removed.

Character.AI moved first and hardest. On 29 October 2025 the company announced it would remove open-ended chat for under-18 users no later than 25 November, ramping teen chat down from an initial two hours a day in the weeks before the cutoff, backed by an in-house age assurance model combined with third-party tools including Persona. It also said it would fund an independent AI Safety Lab. Whatever you make of the motive, that's the largest product in the category deleting its own use case for a whole user segment, and it happened in about four weeks.

California then wrote the rules down. SB 243 requires operators of companion chatbots to give a clear and conspicuous notification that the chatbot is artificially generated and not human where a reasonable person could be misled, to maintain a crisis protocol that refers users to crisis service providers, and, for users known to be minors, to disclose the AI nature of the interaction, remind them to take a break at least every three hours, and prevent sexually explicit material. Operators also have to disclose that companion chatbots may not be suitable for some minors. Annual reporting to the Office of Suicide Prevention starts 1 July 2027. There's a private right of action attached with damages of at least a thousand dollars per violation plus fees, which is the part that will actually change behaviour.

So if an app has recently sprouted an age gate or a "this is an AI" banner, now you know why.

About That Word, Free

Almost every app here has a free mode and almost none of them are free in the sense you mean.

HammerAI's free plan is the most legible, giving you 600 messages a day at an 8,192-token ceiling, and its FAQ says chatting with unlimited characters using local models is free permanently. That's a real free tier because the constraint is stated.

The rest run credits. Nomi sells packs from $4.99 for 40 credits up to $99.99 for 1,800. Character.AI sells Charms in seven denominations from ninety-nine cents to ninety-nine dollars. Kindroid sells selfie credits and audio credits separately from the subscription. Once a product has both a subscription and a consumable currency, the subscription price stops being the price, and I don't think comparing two apps on their monthly figure alone tells you much of anything.

At the shallow end it gets worse. One iOS app ranking on page one for free AI girlfriend searches lists exactly one in-app purchase, a weekly subscription at $4.99. Annualised that's about two hundred and sixty dollars, roughly five times Replika's yearly price, sold as the cheap option. That whole pattern gets its own treatment in the piece on what free AI girlfriend tiers actually hold back.

The Questions That Come Up Every Time

Do any of these actually remember you long term? Some do meaningfully better than others, and I don't believe anyone outside the companies can rank them properly, because only HammerAI publishes the number that would let you. Treat "long-term memory" on a landing page as a feature name, not a measurement.

Is my data safe? Wrong question, I think. The useful question is what specifically happens to the logs, and the four policies I read gave four different answers, ranging from a 60-day post-deletion window to a training archive that outlives your account. Read the retention section before you read the feature list. It takes four minutes and it's the only part of the terms that describes something irreversible.

Are the free ones any good? The genuinely free route is running a local model on your own machine, which costs setup time instead of money and keeps the conversation off anyone's server. Everything else labelled free is a trial with a meter.

What about the ones that promise no filters? They exist, they're adults-only, and the same rules apply to them as to everything else in this article, which is that the pricing should be published, the retention policy should be readable, and if either one is missing that's the finding. I'd hold an unfiltered product to a higher privacy bar than a mainstream one, not a lower one, for reasons I'd hope are obvious.

Will these apps still exist next year? Some won't. Plenty of the names in older roundups now point at parked domains, and when a product shuts down your conversation history usually goes with it. Whether you can export anything is worth asking before you invest a year of evenings. Broader coverage of that risk sits in the companion apps overview.

Does any of this apply to the boyfriend side of the market? Mostly, though the economics there are noticeably rougher, which I've written up separately in the AI boyfriend piece.

Where I'd Start

If I were choosing one today, I'd start from the policy rather than the demo.

Open the privacy policy, find the retention section, and see whether the company will tell you in plain language how long your conversations live and whether anything survives account deletion. Two of the four I read do that clearly. Then check the store listing rather than the marketing page for the actual price, and look for a second currency, because a credit pack next to a subscription means the subscription isn't the number that matters.

Then, honestly, try the free tier of something with a published context ceiling before you pay anyone, because the failure mode in this category isn't that the writing is bad. It's that the thing forgets, four weeks in, and no amount of persona tuning fixes a context window.

I'd also keep Replika's own track record in mind as the base rate rather than the exception, since it's the oldest large player and the one with the most public regulatory history. The newer apps haven't necessarily behaved better. Several of them have simply had less time to accumulate a record.