BabeChat is a Korean-language AI character chat platform where a creator community publishes roleplay characters, and paid ProChat credits unlock named frontier models from Google, Anthropic and OpenAI. Content is split into safety and sensitive tiers under Korea's SafeNet rating framework, with the sensitive tier gated behind carrier-based adult verification.
Used this tool? Rate it
Used this tool? Rate it
BabeChat is a Korean-language AI character chat service in which almost everything you talk to was written by another user. The company's own help centre defines the product as a virtual conversation and entertainment service built on generative AI, and stresses that it delivers immersion not only through text but through images, emotional states and video. That framing matters, because it separates BabeChat from the assistant-style chatbots it superficially resembles. Nobody comes here to summarise a document. People come to inhabit a story that somebody else set up, with a cast somebody else designed.
The operating entity printed in the site footer is Eleven Co., Ltd. (주식회사 일레븐), registered at 338 Gwanggyo Jungang-ro, Building A, Room 828, Suji-gu, Yongin-si, Gyeonggi-do, with Lee Kwan-hyung as representative, business registration number 291-88-02960 and mail-order business filing number 2024-Yongin Giheung-8022. The footer also lists a phone number and a support mailbox. For a category where anonymous shell operators are common, the presence of a Korean mail-order filing number is meaningful: it ties the service to a registered commercial entity that Korean consumers can pursue through ordinary consumer channels.
There is a second layer to the corporate structure, and it is worth stating plainly rather than smoothing over. The Terms of Service, effective 17 June 2026 and labelled as the Republic of Korea country version, name the contracting service provider as BabeChat, Inc., established under the laws of the State of Delaware in the United States and operating under US law. The same document states that this Korean version takes precedence over the BabeChat Global Terms of Service in relation to Korean users. So the structure is a Delaware parent paired with a Korean business registrant, not a single Korean company. Neither document explains the relationship between the two in more detail, and this write-up does not speculate beyond what the documents say.
The site's own metadata describes the proposition as talking freely with your own AI characters twenty-four hours a day and experiencing personalised visual content, under the slogan 한계 없는 AI 채팅, meaning AI chat without limits. Read that slogan against the character creation policy and a more accurate picture emerges: the limits are real, they are written down, and they follow a Korean statutory rating framework. What the slogan is actually promising is a wider expressive range than the sanitised international assistants allow, not the absence of rules.
Because this is a product where content boundaries are the whole story, it is worth knowing which document governs what. The Terms of Service govern the contract, payment and liability. The Youth Protection Policy governs the minimum age and the platform's obligations toward minors. The Character Creation Policy governs what a character may depict and is by far the most specific document of the five. The Community Guidelines govern what may be posted in public community areas, which is a separate question from what a private character may say to you. The Privacy Policy governs data handling, including a delegation table that turns out to be one of the more informative documents on the site.
That five-document structure is not window dressing. Each one is dated, each one is written in operational rather than purely defensive language, and the Character Creation Policy in particular reads as though it was written by somebody who expected to have to defend individual moderation decisions.
The home page after rendering shows six primary tabs — Home, New Release Ranking, Ranking, Category, Tag and Official Creator — and the content beneath them is composed almost entirely of user-authored character cards. Sections such as newly trending characters, popular male-oriented characters, a top ten of popular heroines, and new releases from official creators organise a catalogue rather than a product tour. Cards carry interaction counts, some running into the millions, alongside a creator handle and genre tags covering fantasy, martial arts, action, simulation, romance and much else. Many cards carry a "B ONLY" badge, which marks them as belonging to the platform's own restricted tier.
The practical consequence is that quality varies the way it varies on any user-generated platform. A card with a large interaction count and an active creator who posts update notes in the comments behaves very differently from an abandoned card written in an afternoon. The ranking tabs exist precisely because discovery is the hard problem here.
The Google Play listing, submitted by the developer and displayed publicly by the store, describes four distinct configuration surfaces used to shape a character: prompt, persona, user note and lorebook. The official help centre expands on how they divide the work. A persona carries information about you — the user — so the character reacts to who you are playing. A user note carries information about the character and the situation, and it is also where the automatic summarisation feature deposits condensed conversation history. A lorebook holds world-building entries that surface when relevant keywords appear.
Character authors can set name, appearance, personality and background directly, and the platform provides AI-assisted generation for both the descriptive fields and the images. For image generation the help centre describes entering tags spanning anime styles through to near-photographic looks, receiving four candidate images at once, and revising the tags to regenerate if none of the four work. This is the mechanism behind the visual density of the catalogue: a creator does not need illustration skills to ship a character with a coherent visual identity.
The store description and the help centre agree on the immersion layer. Characters carry emotion images that shift as the conversation moves, chat backgrounds can be configured per room, a story mode presents branching choices rather than free text alone, and a voice mode gives some characters a spoken voice. The help centre adds that a conversation's opening situation can be set and the character's first line rendered as video, which is the feature the marketing copy means when it talks about personalised visual content.
None of this changes what the underlying language model does. It changes how much scaffolding sits around the model's output, and for the roleplay use case that scaffolding is most of the perceived quality.
This is the feature that most distinguishes BabeChat from its immediate peers, and it deserves emphasis because the opposite behaviour is so common in this category. The official help centre publishes a full list of the models available in ProChat mode, names the supplier of each, and prints the credit cost of a single exchange next to it. Gemini 3 Flash costs 31 ProChat, Gemini 3.7 Flash 46, Gemini 2.5 Pro 75, Gemini 3.6 Flash 93, Gemini 3.1 Pro 111, Claude Opus 4.6 and Claude Opus 5.0 both 245, GPT-5.6 SOL 268, and Claude Fable 5 490. Google is credited for the Gemini line, Anthropic for the Claude line and OpenAI for the GPT line.
Aggregators that resell frontier models frequently decline to say whose models they are reselling. BabeChat does the opposite, and it does so in a document that also warns that the list and the consumption figures may change with operating conditions, and that Gemini 3.7 Flash moves from 46 to 67 ProChat from midnight on 23 September. Publishing a price increase in advance in the same table as the current price is a small thing, but it is the kind of small thing that indicates the pricing page is maintained rather than decorative.
Characters divide into two classes under the Character Creation Policy. A safety character is defined as one that any user aged fourteen or over may use, and is permitted content up to SafeNet grade two. A sensitive character is defined explicitly as a character that can only be viewed and used when the safety filter has been switched off. The default state therefore hides the sensitive tier, and reaching it requires deliberate action by the user rather than an accidental click.
The policy publishes the full SafeNet grid it is applying — five grades from zero to four across four axes of exposure, sexual conduct, violence and language — and colour-codes the disposition of each cell. Red means categorically prohibited, orange means conditionally permitted for sensitive characters, green means fully permitted. Grade four exposure and grade four sexual conduct, the categories covering genital exposure and explicit or criminal sexual acts, sit in the red band.
Registration requires a minimum age of fourteen and passes through an initial age confirmation page. The store listing advertises unlimited free chat, with paid ProChat credits reserved for higher-performing models, and the help centre confirms that free mode runs a balanced, safety-oriented conversation model. In practice this means you can browse the catalogue, open a character and hold a conversation without paying, and the paid tier is a quality upgrade rather than an access gate for the base experience.
The sensible first session is to ignore the ranking tabs, open the tag browser, and pick a genre you would actually read a novel in. The catalogue's strength is breadth of premise, and the ranking lists are dominated by a handful of very large cards that are not necessarily representative of what suits you.
Do this before you invest turns rather than after. Fill in the persona with who you are playing — name, gender, and enough background that the character has something to react to. The user note is where you record what you want the character to remember about the situation. Both are per-character on the user note side, so a persona you write once carries across rooms while user notes stay local to the character they belong to.
The reason to do this early is mechanical rather than aesthetic. As will be covered below, long-context retention is the platform's acknowledged weak point, and a well-populated user note is the main lever a user has against it.
The ProChat toggle sits at the bottom of the chat window, and the Pro button opens the model picker. Because each model has a published per-exchange price spanning a range of roughly sixteen to one between the cheapest and the most expensive, model choice is the single largest determinant of what a session costs. A long exploratory session on a Flash-tier model and a short session on a frontier model can cost the same.
Output length is the second lever, and its mechanics deserve attention. All models answer up to 1,500 tokens at the default setting, and only the portion above that guaranteed allowance draws additional credits, calculated in units of 100 tokens. For reasoning-capable models the 1,500 tokens split into 1,300 for the answer and 200 for thinking. Crucially, the deduction is confirmed after the response completes, based on tokens actually used — so raising the multiplier widens the ceiling without guaranteeing that every reply costs that much.
When a conversation grows long enough that accumulated history threatens the model's input limit, the automatic summarisation feature condenses prior conversation into the user note and applies it to the room. This is the platform's official answer to the memory problem, and it is worth invoking deliberately rather than waiting for a failure. Users who care about narrative continuity tend to treat the user note as a living document they edit by hand rather than a field they fill once.
The safety filter can only be switched off by members who have completed adult verification. For users in Korea the path is My → Safety Filter → Adult Verification, and verification is performed using a mobile phone registered in the user's own name. Users living outside Korea must go through customer support for a separate verification process. Until that is done, sensitive characters are neither visible nor usable, and chat histories with adult characters may be hidden.
Credit purchase runs through automatic top-up tiers priced at 4,900 won for 5,000 ProChat, 9,500 for 10,000, 27,500 for 30,000, 44,900 for 50,000 and 89,000 for 100,000, each carrying a first-purchase bonus and a larger bonus after five consecutive purchases. Two subscription products sit alongside: Memory Plus on a thirty-day cycle at 19,800 won against 22,000 won for a single purchase, and an Attendance Check Plus promotion on a seven-day cycle at 2,900 won.
Three billing behaviours are worth knowing before you subscribe rather than after. Completing an automatic top-up setup executes a first payment immediately. All subscription management — joining, changing conditions, changing payment method and cancelling — is web-only, with the app able to display status but not alter it. And a failed payment triggers four automatic retries at six-hour intervals across twenty-four hours before the subscription moves to a paused state and service stops.
The catalogue's most-played cards are not conversation partners so much as campaign settings. Entries with hundreds of images, dozens of named characters, published update logs and explicit notes about which models they were tested against function as serialised interactive fiction. For a reader who wants to be inside a genre rather than merely read it, this is the strongest thing the platform offers.
Because the writing base is Korean light-novel prose rather than translated English, the register is idiomatic in a way that international platforms translated into Korean are not. For an advanced learner this is genuinely useful reading material with a feedback loop attached. For a beginner it is the wrong tool, since nothing in the product is correcting your grammar.
The platform runs creator competitions — the home page carousel advertised a summer creator festival with a fifty million won total prize pool — alongside monthly creator rewards and achievement badges. The Terms of Service establish a formal revenue distribution scheme, though with significant limits covered in the pricing section below.
Because the model picker exposes named models from three suppliers within a single chat room, and re-rolling allows a model change, the platform doubles as an unusually convenient comparison bench for anyone curious about how Gemini, Claude and GPT-class models differ in sustained character work. This is a side effect of the disclosure policy rather than a marketed feature, but it is a real one.
It is not a general assistant, not a writing tool that produces text you own cleanly and outright, and not a private diary. The Terms of Service require users to grant the company a non-exclusive, worldwide, royalty-free, transferable, sublicensable, perpetual and irrevocable licence over derived outputs, and that licence survives deleting the character, deleting the output or closing the account.
The best cards announce themselves. Look for an interaction count in the tens of thousands or higher, an image count stated in the description, a note about which models the author tested against, and comments where the author is still answering questions. Cards that state their tested model environment are telling you something useful: the author has actually checked that their prompt survives contact with a specific model, which is more than most cards can claim.
A Flash-tier model is genuinely adequate for banter, scene-setting and connective tissue. Save the expensive frontier models for turns where you actually need the model to hold complex state or produce something structurally ambitious. Switching mid-conversation is supported, so treating model choice as a per-turn decision rather than a per-session one is both cheaper and better.
Because retention is the platform's weak point, the user note is the only durable memory you control. Rather than waiting for automatic summarisation to fire, write the load-bearing facts into it yourself as the story develops: who knows what, what has already happened, what the character believes about you. When the model does lose the thread, a well-maintained note is the difference between a small correction and starting over.
The multiplier widens the ceiling rather than fixing the cost, so raising it for a climactic scene and lowering it for routine exchanges is a reasonable habit. What is not reasonable is leaving it high permanently and being surprised by the balance. Since deduction reflects tokens actually used, a high setting on a model that answers concisely may cost less than you feared, but the variance is the point.
If your interest extends to the sensitive tier, complete verification first. Chat histories with adult characters can be hidden depending on verification status, so discovering the requirement after investing in a story means an interruption rather than a smooth transition. Verification also determines what appears in the catalogue, which changes what you would have chosen in the first place.
Chat histories are deleted when a creator removes a character or the platform removes it for violations. If a particular story matters to you, copy the passages you care about out of the platform as you go. There is no export feature that will do this for you afterwards.
The clearest fit is a Korean-reading adult who wants long-form interactive genre fiction, is comfortable managing a credit balance, and cares more about narrative range than about assistant-grade reliability. The second clearest fit is a character author who wants an audience — the platform's install base and traffic are large by the standards of Korean-language roleplay services, which is the main reason creators stay.
Users aged fourteen to eighteen occupy a genuinely ambiguous position that this page will not paper over. They may register, they may browse the safety tier, and they cannot switch off the safety filter because carrier-based adult verification will not pass. But the Terms of Service also place minors under nineteen in a separate category for paid services, requiring that the company disclose before contracting that a contract may be voided without legal guardian consent or subsequent ratification. Parents should read that as a design that permits a minor to spend money and then unwind the contract, not as a design that prevents the spending.
International users without a Korean phone number face a materially different product. Adult verification runs through customer support rather than the automated path, the interface and the overwhelming majority of character cards are in Korean, and the creator revenue distribution scheme does not apply to them at all.
Anyone who needs a durable, exportable record of what they wrote should look elsewhere. Chat histories disappear when a creator deletes a character or when a character is removed for policy violations, and there is no user control over either event.
The help centre states that the service is currently available through PC or mobile web browsers and through a beta version of the app. Both stores carry it: the App Store listing shows first release on 28 March 2025 with version 0.1.19 published 21 August 2026 and a minimum requirement of iOS 15.1, while the Google Play listing showed an update date of 24 August 2026.
There is a significant asymmetry between web and app that the official FAQ states outright: app store policy prevents sensitive characters from being displayed in the app's front-facing service. The documented workaround is to start a conversation with a sensitive character on the web, after which it becomes visible and continuable within the app's chat history. The same FAQ notes that web and app prices can differ because of app store commission policy, and the billing guide adds that subscription management is web-only.
The practical reading is that the web is the primary client and the app is a companion. Any user whose interest extends past the safety tier should treat the web as the place where the product actually lives.
The base tier is free and the store listing advertises unlimited free chat, running a safety-oriented model. Paid usage runs on ProChat, defined in the Terms of Service as a point-form currency obtained, purchased or received within the service, used as the means of payment for individual paid services. Purchased ProChat expires after one year in principle unless otherwise stated, free ProChat expires per the terms published for each grant, and ProChat cannot be transferred between members.
Per-exchange costs, as published, run from 31 ProChat for Gemini 3 Flash to 490 for Claude Fable 5, with the figures based on the default 1,500-token output allowance and no discount applied to additional output. Top-up tiers run from 4,900 won for 5,000 ProChat to 89,000 won for 100,000 ProChat, with bonuses that scale with tier and with purchase streak — the top tier carries a first-purchase bonus of 1,100 and a consecutive-purchase bonus of 11,000.
Refund terms are restrictive and clearly written. Unused ProChat is refundable at the amount paid at purchase; used ProChat is not. For automatically charged credits and subscriptions, withdrawal is possible only within seven days and only if the service has not been used at all — a single AI conversation can foreclose cancellation for that billing cycle. Where the company terminates for user fault, it may deduct a refund fee of ten percent or 1,000 won, whichever is greater.
On the earning side, the revenue distribution clause applies only to users resident in Korea and only to revenue defined by operating policy as domestically generated. Whether overseas revenue is shared is left to company policy, and where it is shared, it proceeds in principle through individual contracts. The company also reserves the right to unilaterally discontinue the revenue distribution programme, without affecting revenue already accrued. Community-compiled sources report a creator incentive rate of ten percent and a sponsorship fee of thirty percent, but those figures come from a user-edited wiki rather than official documentation and could not be verified at first hand, so they should be treated as unconfirmed.
Comparison against the international leaders is best done with the numbers side by side rather than with adjectives. A third-party app analytics site placed BabeChat at 4.5 stars from 371 store ratings against Character AI at 4.2 from 1.3 million, PolyBuzz at 4.4 from 695,000 and CHAI at 4.2 from 565,400, with those three each showing over 12.5 million downloads. The rating holds up; the scale does not compare, and it is not trying to.
The honest positioning is that BabeChat competes on Korean-language prose quality, catalogue depth in Korean genre conventions and an expressive range that the international platforms restrict, while the international platforms compete on scale, cross-platform maturity and multilingual reach. A user who reads Korean comfortably and wants the genre catalogue will find nothing equivalent on the international services. A user who does not read Korean has very little reason to choose this over the alternatives.
Within the Korean market itself, community discussion consistently frames the competitive weakness as long-context retention rather than prose quality, naming rival services that hold conversational state better across long sessions. That is a first-hand-unverifiable community judgement, but it aligns with the platform's own documented acknowledgement of the memory limitation.
Three official documents and two app stores give four different answers, and they should be read together rather than reconciled. The Youth Protection Policy and the Terms of Service both set registration at fourteen and over. The Privacy Policy states in the same breath that the company does not collect personal data from children under fourteen and that its websites, applications, products and services are intended for adults — a statement that sits uneasily beside a fourteen-year-old minimum. Both app stores rate the app at 12+, below the platform's own registration floor. And the Terms of Service treat users under nineteen as minors for the purposes of paid services.
At the registration stage, the mechanism is a claim rather than a check: the user asserts that they are fourteen or over, and that assertion is used to confirm eligibility. The Youth Protection Policy states in terms that BabeChat.ai is not responsible for inaccurate or false information regarding user age, and that complying with local laws on content access and consumption is the user's own responsibility. Providing false age information is framed as a Terms violation, which is a consequence for the user rather than a barrier.
It is important not to collapse these two layers. Switching off the safety filter requires completed adult verification, and in Korea that runs through a mobile phone registered in the user's own name — the standard Korean real-name identity route, not a self-declaration. The Privacy Policy's delegation table names NICE Information Service Co., Ltd. as the processor handling identity verification, which is one of Korea's designated identity verification agencies, so there is real infrastructure behind the claim. The Terms of Service separately prohibit circumventing adult verification by VPN or other technical or deceptive means, with usage restrictions for those caught. Publishing an anti-circumvention clause is itself evidence that the gate is treated as load-bearing.
The residual gap is not the verification mechanism but the population it does not cover: a fourteen-year-old can register and browse the safety tier without any check, and overseas users route around the carrier mechanism through a customer support process whose rigour is not documented publicly.
Users arriving with the expectation of an unrestricted adult platform will find that the ceiling is written down. Even within the sensitive tier, the Character Creation Policy prohibits excessive sexual exposure, requires that genitals, pubic hair and the anal region never be directly exposed and that any obscuring be opaque enough that contours cannot be discerned, and categorically prohibits explicit depictions of sexual violence, bestiality, necrophilia, group sex, incest, sadomasochism and voyeurism as non-distributable under Korean information and communications review regulations. Photorealistic images and anything resembling a real person, including deepfakes, are prohibited outright.
Under Korea's Act on the Protection of Children and Youth against Sexual Abuse, the policy prohibits characters that would objectively be recognised as children or adolescents from appearing in sexual contexts. The policy goes further than most in closing a specific loophole: even where a character is explicitly labelled an adult, a school uniform combined with bodily exposure or sexual expression may be classified as child sexual exploitation material and sanctioned under domestic law. The recommendation is to use a uniform that clearly reads as adult, and to avoid school uniforms entirely where sexual elements are present.
The company states directly that because generative models have a maximum input token limit, longer conversations accumulate history until the character may become unable to generate a reply, and offers automatic summarisation into the user note as a partial remedy. Community discussion is blunter, with users circulating a rule of thumb about wrapping up a story within roughly forty to eighty turns and treating manual memory-book maintenance as the alternative. The community figure is unverifiable at first hand and is reported here as community sentiment, but it points in the same direction as the official acknowledgement.
Because credit deduction is calculated on tokens actually used and confirmed after the response completes, a user raising the output multiplier cannot know the exact cost of an exchange before requesting it. Combined with an eightfold range between the cheapest and most expensive model and a documented price increase already scheduled, budgeting requires attention rather than a set-and-forget subscription.
Outputs you derive from a character are stated to be yours, but the character itself, other outputs, other users' content and company-created images and text are not. Against that, you grant the company a perpetual, irrevocable, sublicensable, worldwide licence covering translation, reproduction, display, distribution, modification, derivative works and commercialisation, surviving deletion of the character or output and closure of the account. You may not represent outputs as human-created, nor use them to train your own machine learning models.
The Terms of Service cap total company liability for all damages, losses or causes of action at the greater of what the user actually paid or one hundred US dollars, while noting that some jurisdictions do not permit such limitations. Separately, the terms acknowledge that outputs may not be unique between users, that the service may generate identical or similar outputs for different users, and that outputs may contain AI hallucinations and be inaccurate, objectionable or inappropriate, with the user responsible for evaluating accuracy.
The Privacy Policy discloses overseas transfer to three US recipients: Stripe for payment processing, and Google LLC and Amazon Web Services, both of which receive what the policy describes as the entirety of collected user personal information for cloud server storage. Users may refuse overseas transfer via a published phone number and email, with the policy stating that service provision may be restricted if they do. Statutory retention runs to five years for contract and payment records, three years for consumer complaint and dispute records, and six months for advertising records.
Moderation runs on a proprietary LLM-based content moderation filter, and where that filter detects content violating the terms, the company reserves the right to review the flagged material manually and act on it, including terminating the account. Content assessed as inappropriate may have publication suspended immediately without prior notice, with reinstatement, retention or deletion decided on review. Posts designated as statutory youth-harmful media remain restricted until the government issues a notice lifting the designation. Chat histories are deleted when a creator removes a character or when a character is removed for violations — a real risk for anyone invested in a long story on a card they do not control.
App Store data retrieved directly from Apple's lookup API shows 4.50 from 375 ratings, and the Google Play listing shows 4.7 with a histogram summing to 1,867 ratings across 100,000-plus downloads. Both samples clear the threshold for drawing a quality conclusion and both point the same way. Against that, third-party review aggregation reports persistent complaints about app stability, login errors, reductions in free chat rewards and deepening monetisation, and community discussion describes a pattern of operational incidents and unsatisfying support responses, particularly through 2026. Community-compiled sources also report that a currency restructuring on 27 February 2026 redenominated one ProChat as one hundred ProChat, and that the removal of consecutive attendance bonuses on 1 January 2026 cut thirty-day full-attendance event credits by roughly forty-one percent. Those last figures come from a user-editable wiki and could not be verified first-hand.
The home page returns a shell to non-JavaScript clients, with the visible content injected at runtime. This affects accessibility tooling and archiving more than ordinary browsing, but anyone attempting to preserve a record of a character card should be aware that saving the page source will not capture it.
Yes, in a real sense rather than a trial sense. The store listing advertises unlimited free chat on a safety-oriented model, and you can browse the catalogue and hold conversations without paying. Paid ProChat credits buy access to higher-performing named models, longer outputs and certain convenience features. The free tier is a permanent floor, not a countdown.
Registration requires fourteen or over, and at that stage the mechanism is self-declaration through an age confirmation page — the policy states plainly that the company is not responsible for false age information. Adult verification is a separate and much stronger gate: switching off the safety filter requires verification through a mobile phone registered in your own name in Korea, or through a customer support process if you live abroad. So the fourteen-plus floor is declared, while the adult tier is verified.
It hides sensitive characters, which the Character Creation Policy defines as characters viewable and usable only when the filter is off. Safety characters — everything visible by default — are limited to SafeNet grade two content. The policy publishes the full five-grade grid and colour-codes each cell, with grade four exposure and explicit sexual conduct categorically prohibited even in the sensitive tier.
The help centre names them with prices attached: Gemini 3 Flash, Gemini 3.7 Flash, Gemini 2.5 Pro, Gemini 3.6 Flash and Gemini 3.1 Pro from Google; Claude Opus 4.6, Claude Opus 5.0 and Claude Fable 5 from Anthropic; and GPT-5.6 SOL from OpenAI. Costs per exchange run from 31 to 490 ProChat at the default output setting. The document notes the roster and consumption may change with operating conditions.
Because app store policy prevents them from being displayed in the app's front-facing service, as the official FAQ states. The documented route is to begin the conversation on the web, after which it appears in the app's chat history and can be continued there. The web client is effectively the primary one for anything beyond the safety tier.
Because generative models have a maximum input token limit, and a long conversation accumulates history until the character may be unable to generate a reply. The official remedy is automatic summarisation, which condenses prior conversation into the user note and applies it to the room. Maintaining that user note by hand is what experienced users do rather than relying on the automatic pass alone.
Every model answers up to 1,500 tokens at the default setting with no extra charge, and only the portion beyond that draws additional credits, counted in units of 100 tokens. Reasoning models split the allowance into 1,300 tokens of output and 200 of thinking. The deduction is confirmed after the reply completes based on tokens actually used, so a higher multiplier raises the ceiling without guaranteeing a higher bill.
Only within narrow limits. Unused ProChat is refundable at the price paid; used ProChat is not. For automatic charges and subscriptions, withdrawal requires a request within seven days and complete non-use of the service — one AI conversation can foreclose it. Where the company terminates for user fault, it may deduct the greater of ten percent or 1,000 won as a refund fee.
Outputs you derive are stated to be yours, but not the character itself, other outputs, other users' content or company-created material. You also grant a perpetual, irrevocable, worldwide, sublicensable licence covering reproduction, distribution, modification and commercialisation, which survives deleting the content or closing your account. You may not present outputs as human-written or use them to train your own models.
Processing is delegated to Toss Payments and Stripe for payment, NICE Information Service for identity verification, Channel Corporation for support, Google LLC and Amazon Web Services for cloud storage, and Eleven Co., Ltd. itself for registration, operations and AI service delivery. Stripe, Google and AWS receive data in the United States, with Google and AWS receiving what the policy calls the entirety of collected user personal information. You may refuse overseas transfer, but the policy states service provision may then be restricted.
