Chai App Review: Where the Free Chats Run Out

There are two products called Chai, made by the same company, and the company's own FAQ says they share no data with each other.

That's the first thing a Chai app review has to sort out, so here's the direct answer. The Chai on your phone is the one with 10 million Play installs, ads, and a free tier the company has never put a number on, with Premium at $13.99 a month or $134.99 a year and an Ultra tier at $19.99 a week. The Chai at chai-ai.com is a separate web product whose paid plans carry no ads, with a 24-hour trial that rolls into $159.99 a year, and a metered add-on at $0.66 per 1,000 generated tokens. Neither publishes a memory figure. Both say they fine-tune models on your de-identified conversations and keep data for up to five years after you delete the account. The app is worth its free tier if you can stand an ad roughly every eight to sixteen messages, which is the only frequency I found on any Chai page, and the web version is worth paying for only if you already know you'll use it daily.

Everything below was read on 15 September 2026 from the US App Store and Google Play listings, chai-research.com and its EULA, privacy notice and 2026 investor memo, and the pricing, FAQ, terms and privacy pages at chai-ai.com. The two reviews ranking for this search are usage diaries from 2025. This is the paperwork.

There Are Two Chais, and They Don't Share Data

Chai Research's corporate homepage links to the iOS and Android apps and then says "CHAI WEB Coming Soon". The web product is already live at chai-ai.com, with its own pricing page, terms and support address, and its FAQ describes the relationship in one sentence. "Both Chai Web Pro and the Chai App are products of Chai Research Corp. They are separate platforms with no data sharing between the two."

A subscription on one buys nothing on the other, and the prices don't match either.

Prices from the App Store in-app purchase list, the Play listing and the chai-ai.com pricing and FAQ pages, all read 2026-09-15.

Two rows deserve more than a glance. The web weekly plan at $9.99 is $519.48 a year if you never cancel, against an annual at $159.99, and the app's Ultra weekly at $19.99 is $1,039.48 on the same arithmetic. Weekly plans exist, I assume, because the number looks small, and the App Store review from May objecting to "10 to even 20$ A WEEK" is describing that entry. The second row is CHAI MAX. The FAQ says MAX Auto "removes all rate limits" at $0.66 per 1,000 generated tokens, while MAX Monthly at $90 "doubles your rate limit", which tells you Pro has a limit without telling you what it is. Ninety dollars at the Auto rate buys about 136,000 tokens of output. The FAQ's reason for the web price is that "the AI's response takes 8x more compute than an average app user".

The Play range is worth its own sentence. When this site read the same listing on 29 July for the Play-side waifu app audit, the top in-app purchase was $674.95, already the highest single price this site had found on any companion listing. Today it's $899.99. Apple's list for the same app tops out at $134.99, so whatever that item is, it's Android-only, and neither store says what it buys.

What the Company Says It Built the Model to Do

The chai-research.com homepage is written for engineers and investors rather than users, and I think that makes it the most useful review of Chai there is.

It's an engineering timeline. March 2023, "training a reward model directly on user signals", credited with "a huge boost in our day 30 user retention". May 2023, Proximal Policy Optimisation used to "decrease the probability a chat session ends". June 2023, a larger reward model "trained with 100 million signals to decrease user retry rate and increase chat session length". November 2025, Mixture-of-Experts research "with the goal of increasing user revenue and screentime". December 2025, "Platform screentime and both revenue increased by 25%."

I'm quoting at length because no ranking review does. Chai's model is trained, by the company's own account, on retries, edits and whether the conversation ended, toward the objective of the conversation not ending. The 2026 memo linked from the same homepage lists "User edits (3M / day)", "Message retries (40M / day)" and "End of conversation (3M / day)" among the signals, and claims the result is "1.6 times more retentive and monetizable than OpenAI's GPT-4o-mini". The July 2025 funding note says average session time rose "from 60 to 90 minutes per active session".

If you want a chat that keeps going, this is the product engineered for it. If you're reading reviews because you've noticed you can't put it down, the homepage is telling you that isn't an accident.

The memo also holds the only ad-frequency number I found on any Chai page. On 27 January 2025 the company ran an A/B test on over 20,000 new users, "halving the frequency of ads served to users from 1 in every 8 messages to 1 in every 16", and saw day-30 retention rise 18.5 percent relative, to 19.8 percent absolute. At a hundred messages, that's twelve and a half ads down to six and a quarter. Which setting you're on today, or whether it's a third, isn't published, and the Play review from 12 August describing an ad after "25-30 min" is one user's clock rather than a policy.

The Free Tier Nobody Numbers

Chai does not print a free message allowance. Not on the stores, the corporate site, the EULA, or the web FAQ, which says only that the app is "free to start" and that Pro brings "endless conversations".

The reviews fill the gap with numbers of their own. One says "around 70 messages per day". The other says "70 messages every 2.5 hours". Both are observations from 2025, and the second describes a rolling window rather than a daily cap, which is a different product. I can't tell you which is right today, and I'd rather say that than pick one.

What the company does say is odder. The Safety Center's section on protecting young people includes the line "In some regions, use of the service requires a paid subscription", with no list of regions. A free tier that exists in some countries and not others, stated as a safety measure, is the closest thing to a free-tier policy Chai publishes, and the related search "Chai app is available in which country" is probably people bumping into it.

Then there's what users say happened this year. The Play review from 12 August 2026, with 258 helpful votes, says the app "added back limited messages" and asks "What was the point of taking it away if you were just going to add it back?" A May review on the App Store says "now all of a suddenly ads on the top of chats, limited messages?" So the free tier was uncapped, then capped again, and the cap is whatever your screen shows. My concern isn't the number. It's that a free tier with no published number can change under you without anything the company said becoming false.

What Happened in 2026, According to the Stores

The related search "What happened to Chai AI app" has an answer in the store reviews and version notes.

Between April and August, the most-upvoted reviews on both stores describe three things. Chats erased back to their opening line, "All chats have been erased to the very top of their introduction" on 5 May, "My chats are disappearing, erasing my hours of writing" on 7 June. Messages deleted from bots you haven't opened recently, per the August review. And the return of message limits. The 27 April Play review, with 2,742 helpful votes, is the elegy. "I don't want the voice and images and video. I just want to roleplay and write. It had such amazing memory." Past tense throughout.

The App Store version history corroborates a shake-up without explaining it. Release 2.183 on 31 July says "Updated AI Models". Release 2.184 on 5 August says "Additional features and bringing back bot builder and search", so for some period before August the app shipped without the ability to build a bot or search for one. Every other release from June to the 14 September build is a bug-fix note.

The ratings tell the rest. The App Store has Chai at 4.3 from 278,000 ratings. Google Play has it at 3.6 from 832,000. A ranking review from November 2025 quotes Play at 4.3. The two stores now disagree by seven tenths of a star on the same app, which in my reading is what a rough year looks like from outside.

Filters, on Paper

Chai's reputation is that it has no filter. Its documents say something more specific.

The Safety Center says users "can enable safety filters in the app to filter out mature content in search, hide explicit messages and block specific characters". Enable. The filter is opt-in, which explains the reputation. The same page then says the models "are designed to refuse the generation of restricted material, including sexually explicit content, self-harm content, and any content that sexualises minors", and that every message is scanned for self-harm in real time.

The EULA, last updated 28 July 2026, draws the line more carefully. Section 7 lists six banned categories and cites the Australian Online Safety Act 2021 as the reason. The first is "sexually explicit images or videos of any kind". Section 8 restricts the platform to "users aged 18 and older" and says "you may encounter or generate adult-oriented content, including mature themes and fictionalized violence", provided it "must remain strictly fictional". So the written position is that adult text is expected and explicit images are banned, and the web FAQ's answer on NSFW is that "some user-created content may contain mature themes" under "both automated filters and manual review processes".

Set that against the ranking review's "Not once did I hit a content filter", and both can be true. A text-only product with an opt-in filter and an image ban will feel unfiltered to someone typing prose. Where the Safety Center's "designed to refuse ... sexually explicit content" lands in practice, I can't say from the documents.

Memory, Which Nobody Publishes

No Chai page gives a context length, a memory tier, or a number of messages the model looks back on.

The memo says compute is being used to "extend context windows", which is a direction rather than a figure. The top App Store review, from a blind user praising the app's VoiceOver support, asks for a larger context window in the same breath. The August Play review says bots "reference events from past messages" and, in the same edit, that the app now "deletes bot messages if you havent chatted with them for a while". The two ranking reviews rate memory "Excellent (last 20-30 messages)" and "Bots forget what they said five messages ago", which is the range you'd expect from two people guessing.

I'd put Chai in the same box as most of the category here. The memory is whatever the current model does, the company doesn't say what that is, and the one behaviour users documented this year is deletion.

Who Reads Your Chats

The app's privacy notice is dated 2 July 2026 and framed under GDPR, which is unusual for a Palo Alto company and useful, because GDPR notices have to name a lawful basis for each use.

The one that matters is model improvement. "We use personal data derived from your in-app conversations with chatbots to enhance and fine tune the artificial intelligence models on which these chatbots operate. Before we use any conversation data in this way, we remove personal identifiers." The lawful basis is legitimate interest, so there's no consent box and no opt-out toggle in the notice, and the objection right it describes is offered, in the notice's wording, to citizens of the EU or the UK. Ads are a second use, "if you have a free subscription", with usage patterns shared with advertising partners if you installed from an ad. Retention runs to "the duration of the account and five (5) years following its deletion". Deletion is in the app, under Profile, then Settings, then Delete Account, and you sign in with Google, Facebook or Apple, never a password.

Google's Data safety panel is blunter than the notice. It lists approximate location and your installed apps as shared with third parties for "Advertising or marketing", and in-app messages, search history and user-generated content as collected. The web FAQ, meanwhile, says "Chai AI does not share your personal data with any third parties", which I'd read as marketing shorthand for a notice that names processors and advertising partners.

Then the question in the People Also Ask box, which is whether Chai reads your chats. From the documents, conversations are stored, used for training after identifiers are stripped, and subject to "manual review processes" for moderation. What no document addresses is the top App Store review's claim that "the creators of bots can see people conversations with the bots". I searched the EULA, both privacy notices and the Safety Center and found nothing either way. Treat it as unverified, and don't type anything into a public bot you'd mind its author reading.

One more clause. The EULA's contribution licence grants Chai "an unrestricted, unlimited, irrevocable, perpetual, non-exclusive, transferable, royalty-free, fully-paid, worldwide right" to "sell, resell, publish" what you post, "including, without limitation, your image and voice". It's the standard broad grant, and one of the five Trustpilot reviews is about that clause rather than the app.

Where It Runs, and the 18+ Question

iOS 14 or later, Android, and the separate web product in any browser. The iPhone app is 188 MB and English-only on the store, and the Play listing says character creation is "now open to everyone", with bots kept private or shared.

The age story has a wrinkle. The Safety Center says "The CHAI app is rated 18+ on the Apple App Store and Google Play". Apple's page does show 18+ today. Google Play shows Mature 17+, which is the store's tier rather than Chai's choice. The EULA's own answer to "Why is the Chai app 18+" is section 8, which restricts the platform to adults because "our artificial intelligence models allow for expansive, user-driven roleplay", and has you affirm you've "provided accurate age verification", which as far as the documents show means affirming it. The Safety Center gives Australians an escalation path to the eSafety Commissioner, which together with the Online Safety Act citation in the EULA suggests where the 2026 rewrite came from. The other store-first app in this cluster has a stranger store history, which the Talkie AI review reads from its listings.

The company is Chai Research Corp., a Delaware corporation with a Palo Alto address on the EULA and a Newark, California address on the web terms. The July 2025 funding note puts total investment "over $55M" from Ken Howery, Bill Elmore, CoreWeave and AMD, the memo adds a $30M AMD round and "$85M/yr in revenue in April 2026", and the homepage says "$70M/yr". Two revenue figures on two company pages, and I'd take either as the shape rather than the number.

Questions

Does Chai AI read your chats? Conversations are stored, used to fine-tune models after identifiers are removed, and subject to manual review for moderation. Whether a bot's creator can read your chat with that bot isn't addressed in any Chai document, and a top App Store review says they can.

Is Chai really that good? It's engineered, by its own timeline, to keep sessions going, and the funding note says average session time rose from 60 to 90 minutes. If that's the measure, yes. If the measure is memory or a documented free tier, it publishes neither, and I'd weigh that before the star ratings.

What is the Chai app used for? The EULA's description is "Send messages to AI bots". User-created characters, text only, with an opt-in filter, on a platform restricted to adults.

What's better, c ai or Chai? Chai's own FAQ argues it serves "51,000 unique AI models" against Character AI's one. The practical differences are the opt-in filter, the ads, and the price. The wider field is in the Character AI alternatives comparison, which was written for this question.

Why is the Chai app 18+? Section 8 of the EULA, because the models "allow for expansive, user-driven roleplay". Apple shows 18+. Play shows Mature 17+.

Can I trust Chai AI? It's a funded Delaware company with a GDPR-style notice, a five-year retention clause, a broad content licence, a training clause with no opt-out, and a year of store reviews about erased chats. None of that is hidden. All of it is worth reading before you build anything you'd miss.

What I'd Do Before Paying

Decide which Chai you're buying, because there are two and they don't talk to each other. If your characters and chats are in the app, the web product starts you from nothing.

Then count the ads yourself over a normal session and compare against the two settings the company has published, one in eight and one in sixteen. If you're seeing more than that, the free tier has changed again since the memo.

Then read the plan screen for the word "week". The $19.99 Ultra weekly is over a thousand dollars a year on autopilot, the web weekly is over five hundred, and the annual plans are the only two entries I'd call a price. The site's companion chatbot cost table has Chai's annual at 80 percent of twelve monthlies, beside a competitor whose annual is 29 percent of the same sum.

And if you're going to type anything you'd want back, keep your own copy. The company's documents say it can train on it, keep it for five years after you leave, and delete it from your screen at any time. This year, according to the people who rated it, it did the last one.