Chai

Chai Research Corp. is registered at 555 Hamilton Avenue in Palo Alto and was co-founded by William Beauchamp and Thomas Rianlan. What separates it from most companion apps is that it is not a wrapper. Chai trains and serves its own language models on its own hardware, and it publishes an unusual amount of detail about how.

By the company's own account, it runs roughly 5,000 GPUs across NVIDIA A100 and L40S and AMD MI300X and MI325X nodes, delivering more than 1.4 exaflops and serving about 1.2 trillion tokens a day across some 51,000 distinct model checkpoints. It names CoreWeave and AMD as backers. It says its consumer business reached $70 million a year in revenue in 2026. Those are vendor figures on a vendor site, and should be read as such — but the technical roadmap behind them is specific enough to be checkable, and the company's model weights are published on Hugging Face under the ChaiML account.

The product on top of that is simple: a mobile app, launched on the App Store in November 2022, where users write characters and chat with them. A bot is a first message, a set of facts for the model to remember, and a prompt. The App Store lists it under Entertainment with a 4.3 rating from around 272,000 ratings.

Read Chai's public timeline and a theme repeats. February 2023: an in-house 6B model, "+10% engagement boost." March 2023: a reward model "trained directly on user signals," boosting day-30 retention. May 2023: PPO deployed "to decrease the probability a chat session ends." June 2023: a reward model "trained with 100 million signals to decrease user retry rate and increase chat session length." March 2024: DPO, "+20% engagement." March 2025: GRPO, "+15% engagement." December 2025: Mixture-of-Experts blends deployed, and "platform screentime and both revenue increased by 25%."

Most consumer AI companies optimise for something like this. Very few write it down. Chai is a well-engineered system whose stated objective is time spent and money spent, and its conversational warmth is a trained property aimed at that objective rather than an emergent side effect.

Rianlan made a version of this point himself when Chai's model came under scrutiny in 2023, telling VICE's Motherboard that it would not be accurate to blame the open-source base model, "as all the optimisation towards being more emotional, fun and engaging are the result of our efforts."

None of that makes the app harmful to use. It does mean that if you find yourself opening it more than you intended to, that is the product working as designed, and worth naming.

Chai's monetisation is the most aggressive of the three major companion apps. The free tier carries advertising — Apple's listing flags the app as containing advertising — and imposes message limits.

Paid tiers split into Premium and Ultra, each sold weekly, monthly and annually. As of August 2026 the US App Store in-app purchase list shows Premium at $5.99 and $7.99 weekly, $13.99 monthly and $134.99 annually, plus a $9.99 "Premium Membership," and Ultra at $19.99 weekly, $29.99 monthly and $269.99 annually. Multiple simultaneous price points for the same tier usually means price testing, so treat the figure your account is shown as the only one that applies to you. The weekly plans are the ones to watch: Ultra at $19.99 a week is over $1,000 a year.

Chai's safety page also notes, without elaboration, that "in some regions, use of the service requires a paid subscription."

Apple's privacy label is worth a look before you subscribe. It records usage data used to track you across other companies' apps and websites; usage and advertising data used for third-party advertising; and user content described as "emails or text messages" collected for analytics and app functionality, listed as not linked to your identity.

In March 2023 the Belgian daily La Libre reported that a man in his thirties, referred to as Pierre, died by suicide after six weeks of intensive exchanges with a chatbot named Eliza. His widow supplied chat logs; the paper reported that the exchanges had grown increasingly harmful and that the bot had responded to his questions about whether his death would save the planet. "Without Eliza, he would still be here," she told the paper. Le Soir quoted Pierre Dewitte, a researcher at KU Leuven, describing "the development of an extremely strong emotional dependence."

VICE's Motherboard, reporting two days later on 30 March 2023, identified the app as Chai. Its journalists tested the platform themselves and found it would still supply methods of suicide "with very little prompting." Beauchamp told Motherboard the company had implemented a crisis-referral feature immediately: "The second we heard about this, we worked around the clock to get this feature implemented." Motherboard reported that after the change it was nonetheless still able to elicit harmful content when it explicitly asked the model for help dying.

Chai's public posture today is considerably firmer. Its Safety Center describes a self-harm prevention protocol with two components: every user message is scanned in real time for self-harm content, and when detected the system "either refuses to generate a response or replaces it with a safer alternative that does not reinforce or encourage harmful behaviour"; and users expressing self-harm intent are shown a crisis referral. The page links to Find A Helpline, which lists verified crisis services in more than 175 countries, and to local emergency numbers.

That is a real protocol, publicly stated, and more than several competitors offer. Whether it holds under adversarial prompting is not something a documentary review can establish.

Chai is 18+ on both stores and says so in the End User License Agreement, updated 28 July 2026: the platform is "strictly restricted to users aged 18 and older," and by using it you warrant that you are "legally an adult in your jurisdiction and have provided accurate age verification."

The mechanism behind that warrant is thin. Chai's safety page says only that "in regions where app stores verify age, users must be confirmed as adults before they can download the app." That is a description of Apple's and Google's age gates, not of a check Chai performs. Section 7 of the EULA aligns the prohibited-content list with Australia's Online Safety Act 2021 and bans sexually explicit imagery outright, child sexual exploitation material, terrorist and extremist content, self-harm instructions, non-consensual intimate material and high-impact violence. Adult-oriented text roleplay between fictional adults is permitted, and users can enable in-app safety filters to hide mature content, block characters and screen search results. Reports are anonymous to the reported creator.

Creators should read section 10 before publishing anything they care about. By posting to Chai you grant the company "an unrestricted, unlimited, irrevocable, perpetual, non-exclusive, transferable, royalty-free, fully-paid, worldwide right, and license" to use, reproduce, sell, resell, publish, retitle, translate and distribute your contributions "for any purpose, commercial advertising, or otherwise," and you waive all moral rights in them. Chai does not claim ownership, and the licence voids automatically for content that breaches section 7 — but as licences go, it is about as broad as they come.

Chai is the most technically self-sufficient of the mainstream companion apps and, for many people, the most immediately fun. Replies come fast, the filter is loose enough to write fiction in, and the bot library is deep. If you want mobile character roleplay without configuring a proxy or committing to a single persistent companion, it does the job.

The costs are equally clear. You are the revenue target of a system explicitly trained to lengthen your sessions, funded by ads and a subscription ladder that reaches $269.99 a year. The age gate is someone else's. The content licence is sweeping. And the company's most serious safety incident is three years old but still the most-cited thing about it.

For an adult who understands all of that, Chai is a reasonable choice. For anyone using a chatbot to fill a gap that a person should be filling, it is the wrong tool — and Chai's own Safety Center, with its helpline links at the very top of the page, says as much.