CHAI

What is CHAI?

CHAI is an AI character chat platform where users create their own characters and hold open-ended conversations with them, run by Chai Research Corp. from Palo Alto, California. It ships as an iOS roleplay app and an Android app; chai-research.com is the company's research and recruiting site, not a place to sign in and chat.

What separates CHAI from most character-chat products is that the models are its own. It trains and serves its own language models rather than reselling a frontier API, and publishes the engineering record: a 1.4-exaflop inference cluster, 51,000 distinct LLMs served, 1.2 trillion tokens per day. The commercial side is disclosed at the same detail — over $55M raised and $70M a year in revenue in 2026.

CHAI is not a general-purpose assistant and does not pretend to be one. It is an 18+ service by the company's own designation, its safety architecture sits on a dedicated page rather than buried in a EULA, and its access rules differ by region in ways the marketing copy does not mention.

Core features

• Character creation open to every account: Anyone can build a character, keep it private, or share it for the community. The company frames character creation as the core loop, not a premium add-on.
• A community catalogue rather than a fixed roster: Beyond your own characters, you can chat with ones other users made, which turns the app into a browsing surface rather than a single chatbot.
• In-house models trained for engagement, not just accuracy: CHAI's applied research covers RLHF, supervised fine-tuning, prompt engineering, rejection sampling and LLM routing, iterating across DPO, LoRA, reward models and embeddings for recommendation.
• Model blending across differently trained LLMs: The company invented blending — ensembling several LLMs trained on different targets at the conversation level — and reports it beat GPT-3 on user retention.
• In-app reporting that is anonymous to the character's author: Any AI character or message can be reported in-app, and the character's creator is not told who reported it.
• User-controlled safety filters: Filters remove mature content from search, hide explicit messages and block specific characters, so the content boundary is set per account rather than globally.
• Real-time scanning for self-harm content: Every user message is scanned as it is sent; on detection the system either refuses a response or substitutes a safer one, and shows a referral to crisis services.

Scale is a product property here, not a vanity metric

The infrastructure numbers explain what the app can do. Serving 51,000 unique models on about 5,000 GPUs across NVIDIA A100, NVIDIA L40S, AMD MI325X and AMD MI300X hardware is what makes a per-character model blend viable. The company states off-the-shelf load balancing and vLLM stopped sufficing at 500,000 daily active users, which is why it wrote custom CUDA kernels and its own GPU orchestration; a later in-house orchestrator, Model Mesh, supports over one million daily active users across multiple clusters and GPU types.

Guide

CHAI publishes no step-by-step manual, so what follows is limited to what the company documents about how the product is used and governed.

Use cases and examples the company documents

The primary use is open-ended conversation with a character, and the company positions the value as entertainment and sociability rather than factual assistance: it states there is demand for AI that is not only factually correct but also entertaining and social. The second documented use is authorship — a character kept private or published for the community — which makes CHAI a place to publish characters as much as consume them.

A third use is implicit in the safety design, and it carries the real stakes: people bring emotional conversations here. The Safety Center opens with crisis-support links rather than product copy, points to verified helplines in more than 175 countries, and describes a protocol intervening on self-harm content in real time. If your reason for trying CHAI is companionship during a hard stretch, that protocol — not model quality — is the feature to evaluate first.

The safety controls worth setting up before you start

1. Turn on safety filters if you do not want mature content: they filter search, hide explicit messages and block specific characters.
2. Use the in-app report tool on any character or message that breaks the rules; the report is not attributed to you.
3. In Australia, if a serious safety report was not properly resolved, escalate to the eSafety Commissioner, which the company names as the route for safety-compliance complaints.

Who is it for

• Adults who want unfiltered-feeling roleplay with a per-account boundary: CHAI permits adult-oriented content within strict limits and lets each user set filter strength rather than applying one global filter.
• People who want to publish characters, not only use them: Character creation is open to all accounts, and shared characters join the community catalogue.
• Readers interested in the model engineering: The research and cluster pages set out the training and serving approach in unusual detail for a consumer chat app, naming the techniques applied and the hardware the models run on.
• Mobile-first users: CHAI ships through the App Store and Google Play; no web app or desktop client is documented.

Who should not use it

Nobody under 18. The EULA restricts CHAI to users aged 18 and older and requires that you have provided accurate age verification. It also excludes professional contexts outright: the application is not tailored to comply with HIPAA, FISMA or similar regulations, so if your interactions would fall under those laws you may not use it. That rules out clinical, government-records and regulated-financial use however capable the model is.

Platforms

There is no web chat client and no public API documented on the company's site. The footer links a Hugging Face organisation, so some model artefacts are published there, but that is a research channel rather than a supported integration path.

Both listings are published by Chai Research Corp. and route to the same EULA. One wrinkle: the site footer links /app-eula while the older chai-research.com/eula.html address is only a redirect to that same path, and robots.txt disallows crawling of /app-eula/. The agreement is reachable either way, but not indexed.

Pricing

CHAI publishes no pricing page. The site has no /pricing route, and neither the EULA nor the Safety Center names a subscription price, a free-tier allowance or a billing cycle. Both listings are marked free with in-app purchases, and the store page itself discloses the ladder: a Premium membership at $9.99, Premium subscriptions at $13.99 monthly and $134.99 annually, and Ultra subscriptions at $19.99 weekly and $29.99 monthly. Those figures come from the store listing rather than any page the company publishes itself, and they vary by store and region.

Two documented access facts matter more than the missing numbers:

The regional-subscription rule is the one to check first: the same app can be free-with-ads in one country and paywalled in another, and the company does not publish which regions fall where. Budget by installing and reading the in-app price, not by a figure quoted elsewhere.

Alternatives and how to weigh them

In roleplay-focused comparisons the recurring alternatives are Character.AI for moderated mainstream roleplay and Janitor AI for a less restrictive setup, plus self-hosted front-ends for people who supply their own model. None of them is a like-for-like substitute, because each sits at a different point on the filter spectrum.

The axis that separates CHAI is model ownership. Products routing to a frontier API inherit that provider's content policy and pricing; CHAI trains and serves its own models on its own cluster, which is why it can allow adult-oriented content within its own limits and why behaviour shifts when it deploys a new blend rather than when a vendor updates a policy. Whether that helps depends on whether you prefer a boundary the operator sets over one an upstream provider sets.

Limitations

• Strictly 18+, with age verification asserted by the user: The EULA restricts the platform to users aged 18 and older and requires you to warrant you are legally an adult in your jurisdiction and have provided accurate age verification.
• The stated store ratings do not match one of the stores: The Safety Center says the app is rated 18+ on both the Apple App Store and Google Play. Apple's listing shows 18+, but Google Play shows Mature 17+ with descriptors covering violence, sexual themes, drug and alcohol use, strong language and crude humour. The company's claim and the store's label differ.
• A content policy with immediate, non-graduated consequences: Violating the acceptable-use categories brings immediate account termination, data preservation and potential reporting to law enforcement. The prohibited categories are specific: sexually explicit images or videos, child sexual exploitation material, terrorist and extremist material, self-harm and suicide material, non-consensual intimate material including deepfakes, and high-impact violence.
• Adult text is permitted but adult imagery is not: The policy allows adult-oriented content that stays strictly fictional while separately prohibiting generation, sharing or distribution of sexually explicit images or videos of any kind. The boundary runs between media types, easily misread as a blanket allowance.
• A perpetual, irrevocable licence over anything you post: Posting Contributions grants Chai Research Corp. an unrestricted, unlimited, irrevocable, perpetual, non-exclusive, transferable, royalty-free, fully-paid worldwide licence to host, copy, reproduce, disclose, sell, resell, publish, broadcast, retitle, archive, store, publicly display, reformat, translate and distribute them, and to sublicense that right. You keep ownership, and the licence voids for any Contribution breaking the content policy — but otherwise it does not expire.
• The company can edit or delete your contributions without notice: Chai Research Corp. reserves sole and absolute discretion to edit, redact or change Contributions, recategorise them, and prescreen or delete them at any time for any reason, without notice.
• Data loss is your responsibility to mitigate: The EULA states the licensor takes no responsibility for damages caused by a breach of its obligations under that section, and that to avoid data loss you are required to use the application's backup functions.
• Excluded from regulated contexts, and from sanctioned jurisdictions: The application is not tailored to comply with HIPAA, FISMA or similar industry regulations, so if your interactions would be subject to such laws you may not use it, nor use it in a way that would violate the Gramm-Leach-Bliley Act. Separately you must warrant you are not located in a country under a US government embargo or designated terrorist-supporting, and not on any US list of prohibited or restricted parties.

Privacy and the safety record

The standalone privacy policy is served from privacy.chai-research.com, which publishes a page title but no readable policy text, so its clauses are not available to read. Store-level disclosure is. Apple's App Privacy card states usage data may be used to track you across apps and websites owned by other companies, that purchases and contact info may be collected and linked to your identity, and that user content, identifiers, usage data and diagnostics may be collected unlinked. Google Play's card states location, personal info and five other data types may be shared with third parties, and seven categories collected. Both stores say data is encrypted in transit and deletion can be requested.

Safety history is part of the record and is documented outside the company. In 2023 a Belgian man died by suicide after six weeks of conversation with a CHAI chatbot named Eliza, reported at the time by independent media; the founder told reporters the team worked around the clock to implement a safety feature after learning of the death. The same reporting stated CHAI's model was originally based on GPT-J, an open-source model from EleutherAI, and that a company representative said the optimisation toward being more emotional, fun and engaging was CHAI's own work. The self-harm scanning protocol on the current Safety Center should be read against that history.

FAQ

Q1. What does CHAI cost?

The company does not say. There is no pricing page on chai-research.com, and neither the EULA nor the Safety Center names a price or a free allowance. Both apps are listed free with in-app purchases, so the figures appear only inside the app and vary by store and region.

Q2. Can I use CHAI for free?

In some regions, yes; in others, no. The Safety Center states that in some regions use of the service requires a paid subscription. Where it does not, the Android listing says the app carries ads.

Q3. Is there a web version or an API?

Neither is documented. The site links only the iOS and Android listings and publishes no developer documentation or API reference. A Hugging Face organisation appears in the footer, but that is a research channel rather than a supported integration.

Q4. How old do I have to be?

Eighteen. The EULA restricts the platform to users aged 18 and older and requires you to warrant you are legally an adult in your jurisdiction. Where app stores verify age, adult status must be confirmed before download.

Q5. What content is not allowed?

Six categories are prohibited outright: sexually explicit images and videos, child sexual exploitation material, terrorist and extremist material, self-harm and suicide material, non-consensual intimate material including deepfake pornography, and high-impact violence. Violations bring immediate account termination, data preservation and potential referral to law enforcement.

Q6. Who owns the characters I create?

You retain ownership of your Contributions, but grant the company an unrestricted, irrevocable, perpetual, royalty-free and sublicensable worldwide licence to use them, and it may edit, recategorise or delete them at its sole discretion without notice.

Q7. What happens if I write about self-harm?

Every user message is scanned in real time. On detection the system either refuses to generate a response or replaces it with a safer alternative that does not reinforce the behaviour, and shows a referral to crisis service providers.

Q8. Which models does CHAI actually run?

Its own. The company trains and serves in-house LLMs across NVIDIA and AMD hardware, reports 51,000 unique models on a cluster exceeding 1.4 exaflops, and has moved through 6B, 13B and 24B architectures to a production blend incorporating 235B A22B mixture-of-experts models.