Chai AI Review: The Funniest Bots in the Store, Run Like an Ad Network

My verdict in one line: Chai is worth a free look for silly roleplay on a phone, and worth paying for only once the ads start to grate. The characters are the product, and the best of them write better comedy than most software I'd pay for. Memory, privacy and billing are where it loses points, and it loses plenty.

The numbers behind the app

Chai AI reviews in the two stores read like reviews of two different apps. The App Store gives Chai 4.3 out of 5 from 282K ratings, and it ranks #111 in Entertainment. Google Play gives it 2.2 out of 5 from about 835K reviews, on an Android app with 10M+ downloads.

I read that gap as a story about ads and paywalls more than about the bots themselves. The free tier and friction sections below explain why, and I think they explain it well enough to predict which of the two ratings you'd end up agreeing with.

The money story depends on who is telling it. The chai-research.com site claims $70M a year in revenue in 2026, the roadmap PDF claims $85M a year in April 2026, and Wikipedia reports about $80M ARR. Wikipedia also reports a $2.4B valuation in Q1 2026.

A July 2025 announcement says daily active users grew from 800K to 2M. The roadmap's retention charts show D30 retention of about 19.8%. Roughly one new user in five is still around a month later, and a company that watches that number this closely will treat every ad slot as a dial.

Key takeaways:

• Chai is best for quick, silly roleplay on a phone, with a huge feed of user-made bots to dig through.
• The free tier runs on ads and a message timer, and paying mostly buys quiet.
• Memory is short, so any longer story needs you to repeat the important details.
• The privacy terms are among the loosest I'd sign: chats train the models, and bot creators can see them.
• Pay through the store, and check the store's subscription page before any trial ends.

The rest of this review follows those ten weeks in order. The free tier came first, with its timer and its ads. Then came the step up to Premium, a short and slightly embarrassing detour to Ultra, and one safety message that changed the way the whole app read.

I like that ending, because it is honest. Nothing broke and nobody offended me; the app simply lost its slot once real life had room again. For a companion app I'd call that a fair result, and for a comedy app I'd call it a good one.

A note on what is current. The site's check on 23 September 2026 covers Chai's public pages, its store listings and the Web Lite login page on the web version. No chat was tested and the content settings were not checked in the app, because Chai has no e-mail sign-up.

So the split is simple. How the chats felt comes from early 2024; what the plans cost, what the terms say and how the sign-up works today comes from that September 2026 check. I'd rather tell you that plainly than blend the two into something that sounds fresher than it is.

If you only read one other part, make it the free tier section. That is where Chai decides whether you stay, and it is where the app has changed most since early 2024.

If you arrive expecting a chat window, you get an investor pitch with a data-centre photo instead. That is not a flaw so much as a signal. The company is proud of its cluster and its growth charts, and the chat lives somewhere else.

Web Lite: a black page with three buttons

A minimal web client, "Chai Web Lite", runs at web.chai-research.com. It carries the tagline "CHAT + AI". It is live, but it only offers sign-in with Google, Facebook or Apple.

Web Lite loads instantly. That is easy when there is almost nothing to load: it is a black page with 3 social sign-in buttons and nothing else. There is no guest mode, so there is no way to look around before handing over an account.

A guest mode is the cheapest trust-builder an app has. Without one, the first thing Chai asks for is a link to your main online identity, before it has shown you a single bot.

Behind the login it is a thin companion for browsing the feed and building bots, not a full chat client. I think of it like a proxy file in an edit: good enough to rough things out, not the thing you deliver.

The app is the product

The main product is the iOS and Android app "CHAI: Social AI Platform – Chat". It is maintained often. The iOS app is at version 2.194, updated about 1 day before the 2026-09-23 check, and the Google Play listing was updated on 2026-09-18.

Frequent updates cut both ways. Bugs get fixed quickly, and limits, prices and features can also change from one week to the next, which the free tier section shows in detail.

Chai AI sign-up in 2026

This is where the front door narrows. Web Lite and the app both use only Google, Facebook or Apple sign-in, through Firebase Auth. There is no e-mail or password sign-up option.

The privacy policy says so in plain words: "To access the APP, you need to create a personal account through Google, Facebook, or Apple authentication services." No account could be created during the check, for exactly that reason.

I don't love that design. An e-mail address is the one login most people can make disposable, and social sign-in ties a companion app to the account that holds your photos, your contacts and your work calendar. It is a bit like a bar that only lets you in if you hand over your house keys at the door.

For most people it will be one tap on Apple or Google and done, and I understand why a small team picks the easy route. Still, I'd want the choice. A companion app is the one place where I'd most like my name to stay out of the conversation.

What the first minutes mean for you

In practice the first minutes are short. You install a mobile app of a reasonable size and pick a social login, and whatever onboarding follows sits behind that login. There is no separate web sign-up to compare it against, and no way to try the thing anonymously first.

The gem, for me, was a "grumpy Soviet vending machine" bot. It was honestly one of the best conversations I've had with software, which says more about me than it does about the software. A grumpy appliance with a Cold War accent should not have been the highlight of my winter, and yet.

That is the case for Chai in one character. No studio would green-light a sulking vending machine, and that is exactly why a crowd of amateurs found it.

The other side is just as real. Community creators can be wildly creative, but they can also make bots that are just one catchphrase in a trench coat. For every vending machine there were plenty of characters who had one joke and ran out of road right after it.

The scale has only grown since. The 2026 roadmap memo says 25M+ chat AIs have been created. I'd treat that number the way any editor treats a stock footage library: the size is impressive, and the useful part is still whatever you can find in ten minutes.

The feed decides what you meet first, and a feed tuned for engagement shows what keeps other people swiping. That is not always what you came for, and the odd gems tend to sit a few scrolls down.

Discovery runs through one feed. The Web Lite Home feed is the same "reimagined_feed" the app uses. It comes from production-api-us1…run.app/feed/reimagined_feed, so what you see on the web and on the phone is one list served from one place.

The bot builder has fields for name, image, first message, prompt and memory, plus a Private checkbox. It resizes the bot image to 512px. That is enough for a thumbnail in a feed and nothing more, which is fine, because on Chai the words do the work.

Private by default

The Private checkbox in the bot builder is on by default. I think that is the right default, and a rare one in an app that lives on a public feed. A half-finished character stays yours until you decide it is ready.

Users can list and delete their own bots in Web Lite. Web Lite can also open a /chat/{bot_uid} page for a bot. So the pieces of a small workshop are there: make, keep, delete, open.

The memory field is the one I'd take most seriously. Given how short Chai's memory runs, whatever the bot must never forget belongs there, written plainly, like the slate at the head of a take.

Under the hood

The builder sends model_type "vanilla" and botBuilderVersion V4. Web Lite's public code shows the Firebase project "chai-959f8". Bot deployment posts to bot-service …/bot/config/uploader and is mirrored to Firebase RTDB.

Code comments say Web Lite reuses the same payloads as the Flutter app. In plain English, a bot built on the web should be the same bot on the phone. I'd still check it on the phone before sharing it, the way I'd never send a client a cut without watching the export.

Where it is heading

The roadmap describes "User-controlled SFT", where power users can train the AI, reported at about 2x engagement. It also plans per-agent user-trained AI.

I'm curious and wary at the same time. Letting creators train their own bots could produce more vending machines, which I'd welcome. It could also produce more bots trained on whatever keeps people swiping, and I'd want to see how that is moderated before cheering.

That habit tells you what Chai conversations are good at. They are short, punchy and often funny, and when one misses, the next is a thumb away. I think that is also the trap: an app that makes the reroll this cheap never has to get the first answer right.

The comparison that stuck with me was a shoot with multiple takes. The swipe hands you another take for free, eight deep if you want, and you pick the one that lands. In an edit the chosen take still has to match the scene around it; on Chai nobody checks continuity.

It worked as design and it worked on me. My worry is what it does to a longer story. A slot machine for jokes pays out jokes, not plot.

Why the replies feel uneven

Chai runs its own models in-house. It uses "model blending": an ensemble of several models, chosen per message, described in the paper "Blending Is All You Need" (arXiv 2401.02994). The earlier production mix was 10x24B models, and a Dec 2025 upgrade added 235B-A22B MoE models.

I think that fits the design as much as the bots.

For a comedy bot, that unevenness is almost a feature, because a joke from a new angle is still a joke. For a character you want to know over weeks, it is a problem. Personality is mostly consistency, and consistency is what a per-message ensemble trades away.

The scale is serious. Chai's cluster has 5,000 GPUs (NVIDIA A100/L40S, AMD MI300x/MI325x). It delivers ">1.4 exaflops" and serves 1.2T tokens a day. Chai serves 51K unique LLMs.

I'd read those figures as a statement of priorities. This company spends on hardware and models first, and the modest web client and the thin media side make more sense once you see where the money went.

Trained on what keeps you there

Chai's RLHF runs on engagement signals: thumbs, edits (3M a day), retries (40M a day) and the end of a conversation. The research page lists RLHF, SFT, DPO, LoRA, reward models, rejection sampling and LLM routing.

Those retries are the swipes. Every reroll is a vote, and the models learn from millions of them. I find that clever and a little uncomfortable, because the metric is what keeps people talking, not what is good for them.

There is a fair case for it. A swipe is an honest signal that a reply missed, and no survey gets feedback that fast. The trouble is that engagement also rewards the reply that hooks, flatters or never lets a scene end, and those are not the same as good.

The papers make the goal plain. The research page lists "Rewarding Chatbots for Real-World Engagement with Millions of Users" (2023), "Chai-GPT RLHF Part I: Reward Modelling" (2023) and "Blending Is All You Need" (2024). Talks on the research page include Latent Space (2026) and MLST (2024).

It shows in the numbers. A July 2025 announcement says the average session grew from 60 to 90 minutes. I'd call that a success for the company and a warning label for anyone who opens the app "for five minutes".

I suspect it also explains some of the unevenness, since a reply may come from a model still proving itself in the contest.

Every Chai AI character chat is one-on-one: there is no group or multi-character chat; "each chat is isolated". Every bot lives in its own room. For comedy that is fine, and the vending machine never needed company.

For anyone who wants a party scene or a cast, Chai is the wrong stage.

Who will like the conversations

If your idea of a good chat is a fast back-and-forth with a strange character, Chai delivers, and the swipe makes the bad replies cheap. If you want a slow burn with callbacks and a character who grows, the design pushes against you at every level. The reroll, the blending and the engagement training all favour the next line over the last hundred.

My overall read: Chai conversations are a stand-up set, not a novel. Come for the one-liners and the odd inspired character, and leave the long arcs to apps that take memory and consistency more seriously.

In editing terms that is a continuity error, the kind where the coffee cup changes hands between shots. It was funny, and it was also the moment I stopped expecting Chai to hold a thread. Being asked about a truck I don't own, by a friend who doesn't exist, puts the joke squarely on me.

Other reports are harsher. One review site reports that memory forgets established details within three to five messages. A 7-day memory test at aigirlfriend.coach found the AI invented a "beach trip" that never happened in the conversation.

A user review relayed by another site says the memory forgets a lot.

The worse problem is not forgetting inside a chat, it is losing the chat. Recent 2026 Google Play reviews say chats and bot messages get deleted when inactive, with lines like "more than half my chats are gone" and "All chats have been erased". A companion that forgets your truck is a sitcom; a companion that loses the whole history is a hard drive failure.

My advice is simple. Put the facts that matter in the bot's memory field when you build or pick it, repeat the key details every so often, and screenshot anything you'd hate to lose. Treat Chai as a place for scenes, not for a long relationship with a paper trail.

In fairness, a short memory suits the way Chai wants to be used. A bot you visit for five minutes of jokes does not need your biography, and sitcom characters rarely remember last week's episode either.

If memory is the reason you want a companion app at all, I'd look elsewhere. Chai is built around the next reply, and the next reply is exactly what it is good at.

There is movement on voice. A text-to-speech voice output for bot replies is rolling out on select bots in 2026, reportedly for Ultra only, according to secondary sources. I'd wait for it to reach the store listing before paying for it, because "reportedly for the top tier" is the phrase that sells upgrades.

Images exist, in a small way. A hands-on purchase test by aigirlfriend.coach found 3 free images a day on the paid tier. The same source describes image quality as accurate for anime-style output.

That is a fair niche, and a narrow one.

Video is a blank. No official or secondary source, out of 5+ reviewed, mentions video generation for Chai.

The contrast with the rest of the company is odd. A firm that runs its own models at this scale has put very little of that muscle into pictures or sound for users. I don't count that as a loss, because Chai's characters work on their lines, and a good line does not need a face.

My read: pay for Chai for the conversations or not at all. Voice is on the way for the top tier, images are a small daily extra, and video does not exist, so anyone shopping for a visual or spoken companion should keep scrolling.

The timer changed how the app fit into my day, and not for the better. Instead of opening Chai when I felt like talking, I opened it when the timer said I could, which turned a hobby into a scheduling problem.

There were ads between messages. At one point I watched a 30-second mobile game ad to unlock more chat and thought, "I'm paying for a pretend conversation with my attention span." That line still sums up the free tier for me better than any feature table.

The ad itself was not the problem; the exchange rate was, thirty seconds of a game I would never install for a few more minutes with a bot.

Free Chai was usable, but it was never free in the way a library is free. It was free the way a casino buffet is free: the price is somewhere else, and it is you.

The numbers people quote vary. Secondary sources most often cite a rolling window of about 70 messages per 2.5 hours on the free tier.

A third-party competitor blog puts the free allowance at about 70 messages per batch, which is not verified. That matches my own count from 2024.

Other reports are tighter. A hands-on report at aigirlfriend.coach found about 11 messages before the free chat locked. An April 2026 report describes a free-tier restriction to off-peak hours only.

So my old figure of about 70 is a best case today, not a promise. Depending on the day, the region and the source, free ranges from generous to stingy, and I'd plan for the stingy end.

The ads, by design

Google Play lists "Contains ads" and "In-app purchases". The privacy policy says advertising is shown to users "if you have a free subscription". Third-party sources, not verified in-app, say the free tier has ads, a message recharge timer and short memory.

Here is the part I find most revealing. The public roadmap PDF openly treats ads as a retention lever. In a Jan 2025 A/B test, halving ads from 1 per 8 messages to 1 per 16 raised D30 retention by 18.5% relative.

The free tier therefore shows about 1 ad per 8–16 messages.

I respect the honesty and dislike the result. The company has measured that fewer ads keep more people, and still runs an ad every handful of messages, because the people who leave are not the point. The people who pay to make the ads stop are.

What users report now

Recent 2026 Google Play reviews repeatedly mention heavy ads: one ad after about 25 to 30 minutes, or ads mid-chat. Reviews from Jul/Aug 2026 mention "refresh for free chats" prompts that show ads. That is the same trade I made in 2024, still on the menu.

Free is not free everywhere

The Safety Center says: "In some regions, use of the service requires a paid subscription." Wikipedia says that in Jan 2026 Chai added country-based tiers, and users in "low tier" countries must subscribe. It also says a "basic" tier with unlimited messages plus ads was then announced.

My advice for free users: treat it as a tasting menu. Find two or three bots you like, learn the rhythm of the timer, and decide within a week whether the ads bother you enough to pay. If they do, the next section is about what paying buys.

For browsing and a few laughs, free is enough; for anything longer than a coffee break, the timer and the ads will run your evening for you.

The price has not moved. Premium Subscription (Monthly) still costs $13.99 on the US App Store. I paid $13.99 in early 2024, and the same monthly plan costs $13.99 today.

Ultra, one cycle

I briefly tried Ultra at $29.99 for the better model. It was a little sharper, but not twice as sharp. I downgraded after one cycle and felt like I'd returned a jacket.

Secondary sources say Ultra adds faster, priority responses.

The price list in September 2026

The prices here come from the US App Store in-app purchase list, in USD. The App Store prices in the record were checked on 2026-09-22. Web Lite has no pricing or store section, so the store is the only place to see them.

Premium Membership costs $9.99. Premium Subscription (Annual), also listed as Premium Annual, costs $134.99, which works out to about $11.25 a month. Ultra Subscription (Weekly) costs $19.99, which works out to about $86.62 a month.

The weekly Ultra plan is the one to stare at. Nearly three times the monthly Ultra price for the same tier is the kind of math that only works on people who never do math. I'd skip it outright.

The money math

Here is the math as I'd do it at the kitchen table. Monthly Premium at $13.99 against the annual plan at about $11.25 a month saves a few dollars a month, in exchange for a year paid up front. That is a gym-membership deal: good value for the people who show up, a donation for everyone else.

Monthly Ultra at $29.99 costs more than twice monthly Premium. For me it was a little sharper, not twice as sharp, and I'd need a clear list of what the top tier adds before paying double again.

The $9.99 Premium Membership is the one I'd ask about before buying. It is cheaper than the monthly subscription, and I'd want to know what it leaves out and how often it bills before choosing it over the plan I know.

What each tier adds

Third-party sources, not verified in-app, say Premium removes the timer and the ads. Secondary sources say Premium runs the standard AI model at standard response speed. That matches what Premium did for me: the waiting stopped and the ads vanished.

Is the step up worth it?

For me Premium was worth it and Ultra was not. If the ads and the timer are the only complaint, monthly Premium fixes both. The annual plan saves a little, but only if you are sure you will still want Chai in a year, and most people I'd describe as sure are wrong.

My rule: start on monthly Premium, set a reminder, and never let a trial roll into the annual plan by accident. The cancel section explains why that rule exists.

What Pieces buy is less clear. Coin packs are sold separately from the subscriptions, and no page says what coins unlock. I'd compare it to a gift card for a shop that won't show you its shelves.

The currency is spreading into features. Google Play reviews from Jul/Aug 2026 mention a coin requirement for some creation features. A 2026 iOS review complains about the new coins requirement.

A September 2026 report describes a new "pieces" currency in which every message costs at least 1 piece, with about 200-300 free pieces a day. If that holds, the free tier stops being a timer and becomes a meter, which is a different deal.

In a report's words, "every message costs pieces, one piece at the cheapest, with free pieces handed out daily". A daily allowance of a few hundred sounds generous until each reply has a price tag, and the swipe for a better answer starts to cost.

So there are two meters and a subscription, and they do not share a wallet. It is the vending machine problem in reverse: plenty of coins on sale, and no label on the buttons. Subscriptions run through the store, Pieces are a separate balance, and image purchases go straight to the card.

My advice: don't buy Pieces until the app tells you, on the purchase screen, exactly what they unlock. A $2.79 pack is small, but small unexplained purchases are how a monthly bill quietly doubles.

That is almost refreshing. Plenty of companion apps run a countdown timer on a "limited" discount that resets every time you reload the page. Chai skips the theatre and just sells through the store at list price.

It also means there is nothing to wait for: buying on day one costs the same as buying in a month, so the only reason to delay is to be sure you want it.

The one blind spot is e-mail. No marketing e-mails could be checked, because no account was created. Some apps save their real offers for the inbox or the cancel screen, and I can't rule that out here.

My advice is not to wait for a coupon. If you want Premium, the monthly plan at list price is the deal, and the only real saving is the annual plan, which I'd buy only after a few months of knowing you'll stay.

If a discount does show up inside the app, read which plan it applies to before you tap. A trial pop-up is still a subscription with a start date, and the store will hold you to it.

It made me stop and think about how many people use this thing alone at 3 a.m. I was fine, just tired and cranky about a timeline. Somebody else, typing the same words at the same hour, might not be, and the app had no way to tell which of the two it was talking to.

I learned later that Chai added those safety messages after a widely reported 2023 case in Belgium. Knowing that changed how I saw the whole app. The funny vending machine and the crisis line suddenly lived in the same building, and I stopped treating the place as only a toy.

The case behind the safety text

Chai's bot "Eliza" was linked to a 2023 suicide case in Belgium, which Belgian authorities investigated, per Wikipedia. Crisis-help text was added after the Belgian case.

The Safety Center says every user message is scanned in real time for self-harm. Self-harm messages trigger crisis-resource notifications. I think that is the most important feature in the app, and the one nobody mentions in a store review.

I don't think an app can be blamed for everything that happens in a user's life, and this case should never be a punchline. But I respect a company that changed the product after something this serious, and I'd want every app in this category to carry the same net.

Who the app is for, on paper

The EULA restricts Chai to users aged 18 and older. It also warns that users may meet adult-oriented content, including mature themes and fictionalized violence. Chai is rated adults only on both the App Store and Google Play.

The Safety Center states that CHAI is a service for adults only. It says that in regions where app stores run an age check, a user's age is confirmed before download. It names no other method.

The Safety Center says the models "refuse the generation of restricted material", including sexual content. Users can report characters and messages anonymously. The Safety Center has an escalation section for Australia's eSafety Commissioner.

What users say, and the mixed signals

Is Chai AI NSFW? The signals are mixed. The EULA allows "adult-oriented content, including mature themes and fictionalized violence". Google Play rates the app for "Extreme Violence, Strong Language".

Apple's age rating lists Frequent "Mature or Suggestive Themes" and "Profanity or Crude Humor". It lists Infrequent sexual content, "Horror/Fear", "Cartoon or Fantasy Violence", "Alcohol/Tobacco/Drug" and "Guns".

Meanwhile users looking for spicy Chai AI roleplay praise the lack of filters. Google Play reviewers praise that there are "NO filters". So the rulebook says one thing, the reviews say another, and the truth is probably whatever the bot in front of you was trained on.

My guess is that both descriptions are true in their own place. The written rules set the outer walls, and inside them the community bots roam further than the policy language suggests.

The optional filter

The app has an opt-in safety filter. Optional in-app "safety filters" can filter mature content in search, hide graphic messages and block characters. I'd switch them on for the first week and loosen them once you know the feed.

Chai's FAQ describes moderation as reactive: manual and automated detection, with violating content removed after the fact. In its own words, "some user-created content may contain mature themes". That is a fair warning, and I'd take it literally.

What this means for you

If you want firm boundaries, turn the filter on and report what slips through. If you want loose ones, remember that the terms still apply, and the penalties in the terms section are real.

My view: the filters are loose, the rules are strict on paper, and the self-harm safety net is the part Chai gets most right.

6527981, and the EULA is governed by California law.

In plain English, a US company in Palo Alto writes the rules, and California law decides disputes. For a US user that is ordinary; for everyone else it means the fine print was written far from home.

Refunds and the store

Chai has no refund policy of its own. The EULA points refunds to the store operator, Apple or Google. It refunds the purchase price if the app fails to conform to its warranty, a standard Apple EULA clause: "Your Licensed Application purchase price will be refunded to You".

That clause sounds better than it is. It covers an app that does not work as promised, not a subscription you forgot to cancel. For those, you are at the store's mercy, and I'd plan as if there is no refund at all.

What you give up when you create

Users "retain full ownership" of their Contributions. They also grant Chai a broad licence to them: "unrestricted, unlimited, irrevocable, perpetual ... transferable, royalty-free ... worldwide". That licence includes sublicensing and "commercial advertising".

I'd call that ownership in name only. You keep the deed, and Chai gets to live in the house forever, rent it out and put it in an ad. If a character you build is something you care about, keep the original notes somewhere else.

Breaking the banned-content rules leads to "immediate account termination, data preservation, and potential reporting to relevant law enforcement".

Your chats and your data

The privacy policy is at chai.ml/privacy, updated July 2, 2026, and written in GDPR/UK GDPR style. Chats are used to train and fine-tune the AI models, with personal identifiers removed first. The lawful basis for that training is legitimate interest.

Personal data is used for ads on the free tier. Usage patterns are shared with ad partners for attribution. The App Store privacy label says "Usage Data" is used to track users across other companies' apps and websites.

Data is kept for the life of the account plus up to 5 years after deletion. That is a long tail: whatever you type today may outlive your interest in the app by years. An App Store review points out that bot creators can see users' conversations with their bots, and that this is not disclosed.

That last point is the one I'd tattoo on the download button. Talking to a community bot may mean talking in front of its creator. I'd never type anything into a public bot that I would not say to a stranger in a diner.

Who to contact

The DMCA agent is reachable at [email protected] and (408) 320-7968. Support is reachable at [email protected].

An e-mail address for support and a legal contact for copyright is thin for an app this size, but at least both are published. I'd keep a screenshot of any billing problem before writing to either, since the store, not Chai, handles the money.

Social-only login, with no e-mail option, blocks account creation for anyone without those accounts. No chat was possible during the check: the web client needs a Google, Facebook or Apple account, and the product is otherwise mobile-only.

I think this is the single most avoidable friction in the app. A plain e-mail sign-up and a working web chat are not exotic features. They are the door and the doorbell.

Prices you can't see

Pricing is not published on the website; only the store in-app purchase list shows prices. The difference between "Premium Membership" at $9.99 and "Premium Subscription" at $13.99 is unexplained. Tier features could not be verified.

In my trade, a quote without a line-item breakdown is how arguments start. Two Premium products at two prices, with no page saying how they differ, is that quote.

A store list is not a pricing page. It gives names and numbers, not what each plan includes, which leaves buyers to piece the tiers together from third-party reviews.

Pressure and surprises

Recent 2026 Google Play reviews mention pressure to subscribe. Wikipedia and PiunikaWeb reported in Feb 2026 that unannounced token limits froze chats for free and paid users. Paying customers hitting a wall they were never told about is the fastest way I know to lose goodwill.

In 2026 users complain about chat deletion, ads, token limits and region paywalls. That list reads like a punch card of everything that can go wrong with a subscription app.

Put together, the pattern is a free tier that keeps tightening and a paid tier whose limits can still move. I'd not call that a scam. I'd call it an app that changes the rules faster than it explains them, and that is its own kind of cost.

Nowhere to check

Chai has no official status page; chai-research.statuspage.io is unclaimed. When chats freeze, there is no official place to learn whether the problem is yours or theirs, so working out how to fix Chai AI starts with guesswork.

My advice: screenshot what you care about, keep the store's subscription page bookmarked, and assume any limit can change without notice. It is not a way to enjoy an app, but it is the way to use this one without surprises.

It is also best for creators who want quick no-code bot building (name, prompt, first message, memory) with private or public sharing. If you like writing characters more than living with them, the builder is the most pleasant part of the product.

Now the skip list, which is longer. Chai is not for desktop or web users: web is "Coming Soon" and Lite needs social login. It is not for anyone who won't use Google, Facebook or Apple login.

It is not for privacy-minded users: chats train the models, ads track users, creators can see chats, and retention runs 5 years. It is not for people who need persistent chat history, given reports of chats wiped after inactivity.

It is not for users in "low-tier" countries where a subscription is required. And it is not for anyone who wants an ad-free free tier.

If you sit between those lists, a light user who wants a laugh on the bus and does not care much about privacy, the free tier is a cheap way to find out which side you are on.

I'd call that the most useful thing I learned on Chai. A trial is a subscription with a delayed start, and the delay is where people forget.

How it works today

Payment runs through Apple App Store or Google Play in-app purchase. Auto-renew and cancellation run through the store's subscription settings. No Chai-specific cancellation text was found.

So there is no cancel button inside Chai to hunt for. On an iPhone the switch lives in Apple's subscription list; on Android it lives in Google Play. That is simpler than many apps, as long as you know to look there.

The store also keeps every plan and its renewal date in one list, which is exactly where my near miss got caught.

Deleting the account

The account can be deleted in the app via Profile > Settings > Delete Account. Deleting the account does not cancel a store subscription, so I'd cancel in the store first and delete second.

Remember the retention rule from the terms section as well. Deletion stops the account, not the clock on the data the company keeps afterwards.

If you do choose the annual plan on purpose, set the reminder weeks before the year ends, not days. An annual renewal is big enough to hurt, and a store refund is never guaranteed.

On the free tier

Treat the timer as a feature, not a punishment. Short sessions suit Chai's strengths, and the cooldown stops a five-minute break from becoming a lost evening. If the ads start to annoy you more than the bots amuse you, that is the signal to try one month of Premium, not a year.

Test a bot in a handful of messages. The good ones show personality immediately, and the one-note ones repeat themselves just as fast. Swiping for a better reply is fun, but I'd set a limit, or you'll spend the evening choosing jokes instead of reading them.

Keeping what matters

Put important facts in the bot's memory field and repeat them now and then. Screenshot any conversation you'd hate to lose, since chats have been reported deleted after inactivity. Keep your own notes for any character you build, because the licence you grant Chai lasts forever.

Money

Buy monthly, never weekly. Set a reminder before any trial ends and check the store's subscription list the same day. If a month of Premium does not change how often you open the app, cancel it; the ads were never the real problem.

Leave Pieces alone until the purchase screen says exactly what they unlock.

Its YouTube channel is youtube.com/@Chai-research. Its Hugging Face page is huggingface.co/ChaiML. Its Facebook page is facebook.com/chairesearch, and its Instagram is instagram.com/chairesearch.

The iOS app is at "apps.apple.com/us/app/chai-social-ai-platform-chat/id1544750895". The Android app is at "play.google.com/store/apps/details?id=com.Beauchamp.Messenger.external". Chai's Wikipedia article is en.wikipedia.org/wiki/Chai_AI. The site links no Discord or TikTok account.

I'd install from those two store links only. With a company site that has no chat and a web app still coming soon, the stores are the only real front door.

Memory is weak. Characters lose details within a scene, and reports of wiped chats make it worse than simple forgetting.

The free tier is mixed. It works, and it is generous by some accounts, but the ads, the timer and the shifting rules make it a free tier you manage rather than enjoy.

Images and voice are weak, because Chai is a text app with a few extras bolted on.

Privacy and terms score poorly. Training on chats, ad tracking, visible chats for bot creators and long data retention add up to terms I'd read twice before signing.

Cancelling and billing are good, because the stores handle them, with one warning about trial pop-ups. Content settings are mixed: loose in practice, strict on paper, with a self-harm safety net I respect.

That is both its charm and its problem. The basics any paid app owes its users are the things Chai still lacks: e-mail sign-up, a real web client, clearer price pages and memory that holds up past a single scene. Chai kept the crowd and the chaos instead.

Where Chai wins, in my view, is surprise. A catalogue built by a crowd of strangers produces characters no product team would sign off on, and the best of them are funnier than anything a polished studio app would ship. If you want a companion with a stable personality, a studio app is the safer bet; if you want to be surprised on a Tuesday, Chai still has the edge.

One more difference is easy to misread. Chaiverse is a model-training competition platform, not a consumer or developer API. So Chai is not a toolkit for builders who want to plug a model into their own app; it is a closed app that invites outside models to compete inside it.

The table below puts Chai next to the others, with periods and cheapest plans, so you can set them side by side yourself.

More on this site: apps like Chai AI, Chai vs Character AI, what happened to Chai AI, Chai vs PolyBuzz.

The leftover is small and a little silly: a grumpy vending machine with better timing than most people I know. For an app whose paid tier mostly sells silence, that is a decent leftover. For a company that measures everything, it is probably the one number that never showed up on a chart.