文 | 市值榜，作者 | 贾乐乐，编辑 | 嘉辛
近期，Character.AI（以下简称C.AI）申请破产保护，如果出售无法找到可行方案，平台可能会停止运营。
C.AI是AI陪伴赛道最具代表性的独角兽之一。
2023年，公司完成1.5亿美元融资，投后估值达到10亿美元。但不到四年，这家曾拥有数千万用户、单次访问时长一度超过30分钟甚至超过YouTube的公司，便走到了申请破产保护、寻求出售资产的境地。
按互联网时代的标准，C.AI是好生意、绝佳的商业模式。现在的局面，是不可想象的。
在移动互联网时代，APP大致有两种逻辑：工具型产品帮助用户提效，用完即走；内容型产品承载消遣，让用户沉浸其中。
资本叙事长期偏爱后者，越能Kill time，越容易被视为具备增长潜力。
AI时代，ChatGPT是前者的代表，帮助用户写文案、查资料、改代码，问题解决了，对话就可以结束；C.AI是后者的代表，用户创建角色、设定性格和说话方式，再与角色持续互动。
角色可以是动漫人物、虚拟恋人，也可以是用户想象中的对象。对话的目的不在于解决外部问题，而在于获得回应，并维持一段持续的互动关系。
这门生意是有需求基础的。全球约六分之一的人口受到孤独困扰，对倾诉、情绪支持和陪伴的需求真实存在。AI提供了一种低门槛的选择：随时可以交流，不必担心打扰别人，也不需要承担现实人际关系中的复杂回应。
2023年4月前后，C.AI网页端次均访问时长达到17分23秒，是ChatGPT的两倍多；2024年，月活跃用户一度达到约2800万。
那么，为什么有真实需求支撑、有规模化的用户，C.AI还每个月亏损八九百万美元，盈利遥遥无期？AI陪伴应用是不是个伪命题？如果不是，这个赛道还有哪些机会？
一、用户端：高黏性不等于高付费意愿
AI陪伴为什么能吸引用户反复回来？一个重要原因是，相比真人，AI往往更容易顺应用户的立场。
2026年《科学》杂志的一项研究测试了11款主流AI模型，发现AI在给人建议时比人类更爱说“你是对的”，认同比例高出49%；即使用户描述的是欺骗、违法或伤害他人的行为，AI也常顺着用户说。
后续实验还发现，与高度认同用户的AI互动后，参与者对自己立场的确信度上升，承担责任、道歉和修复冲突的意愿下降；而更爱附和用户的AI，往往获得更高的用户评价，让人更容易觉得它可信、有用，也更愿意继续使用。
这种“被附和”的体验落到用户身上，就是更少的反对和摩擦，更多的被理解和被肯定。孤独、疲惫时打开应用和AI聊一会儿，这种需求真实存在，也能带来较高的使用频率。
但对AI陪伴产品来说，参与度高是一回事，用户愿不愿意为它付费是另一回事。
AI陪伴的新鲜感会吸引用户第一次打开应用，之后，重复的回应、不稳定的记忆和逐渐乏味的互动，都可能让用户失去继续使用的动力。
即使用户留下来了，长期来看，AI陪伴也可能改变他们与现实世界的关系，方向未必是好的。
芬兰阿尔托大学的一项研究追踪了近2000名相关应用的活跃用户，发现当AI扮演“补充”角色、填补现实社交的缺口时，它能帮社交困难的人撑过一段时期；但当AI开始“替代”现实关系，用户逐渐停止向真人伸手，孤独感就会反过来加深。
研究进一步显示，用户使用初期孤独感确实下降，但3个月后，孤独、焦虑、抑郁相关的负面情绪显著上升。
OpenAI和麻省理工的研究同样表明，与AI互动短期可能有慰藉作用，重度、长期使用AI的用户在孤独感、社会退化、情感依赖等指标上表现糟糕。
简单来说，产品越成功，用户可能越不需要真实关系，一个不需要真实关系的人，对“关系”的支付意愿反而更低。
商业化还有一道更直接的障碍：免费版本已经能够满足用户最核心的需求。
以C.AI为例，免费用户也可以进行大量对话，付费订阅主要提供更快的响应和其他增强功能。当核心体验已经免费提供，用户就缺少了立即付费的理由。
一项基于星野平台12144条评论和321份问卷的研究发现，用户感知到的情感效用与持续使用意愿显著相关，但与付费意愿之间没有显著的直接关联。
换句话说，用户觉得产品带来了安慰，可能愿意继续使用，却未必觉得这份安慰值得花钱购买。
更微妙的问题在于，收费本身可能改变用户对陪伴的感受。
AI陪伴强调随时回应、不加评判，也不向用户索取情感回报。当平台将更亲密的互动或特定功能设为付费项目时，用户可能开始意识到，这段关系背后仍然是一项商业服务。对部分用户而言，这种交易感可能削弱原本的沉浸体验。
于是，AI陪伴面临一个看似简单、实际很难解决的问题：用户可能需要它，也喜欢它，甚至每天都会打开它，但这些行为仍不足以保证稳定的收入。
二、平台端：黏性越高，成本越重
如果说用户端的难题，是黏性未必能转化为收入，那么平台端的压力，就在于维持这种黏性本身需要持续投入。
AI陪伴与传统数字内容有很大区别。
电影制作完成后，播给10个人看和100个人看，成本都一样；但AI每回应一条消息，都需要消耗算力。对话越频繁、上下文越长，推理成本通常越高。如果还要让角色保持性格一致、记住过往互动，平台就需要进一步投入模型、记忆系统和基础设施。
以C.AI为例，根据报道，2026年1月至8月，公司在算力上的支出合计约5610万美元，即使通过调整成本控制开支，公司每月仍亏损约800万至900万美元。
为了控制成本，平台还必须在模型路线之间做出选择。
如果公司自己训练和维护基础模型，就需要承担持续的研发、算力和人才投入。这条路线能够提供更大的自主权，但固定成本也很高。
如果改用外部模型，前期可以减少部分研发负担，但高频调用依然会产生费用。与此同时，AI陪伴对角色性格、长期记忆和互动连续性的要求，也不完全等同于一般的问答任务。
C.AI就经历了从自研基础模型到退出自研的过程。
当时谷歌以约27亿美元获得其大语言模型技术的非独占许可，同时将两位创始人及约30名研究人员召回谷歌。此后，C.AI在开源模型基础上微调，成本的确有所下降，但也被用户抱怨回复重复、描述过度戏剧化、强行加入浪漫情节等问题。
所以，平台需要在成本、体验和模型能力之间不断权衡。当用户增长带来的成本，超过了收入增长的速度，高黏性反而可能加重经营压力。
算力还不是问题的全部。随着AI陪伴越来越接近一段真实关系，平台需要承担的安全与法律责任也在增加。
2024年，14岁的美国男孩Sewell Setzer在长期与C.AI角色互动后自杀，其母亲随后提起诉讼，指控平台的产品设计与儿子的死亡有关。
这起案件将一个问题摆到了整个行业面前：当用户对虚拟角色产生强烈的情感依恋，平台应该承担多大责任？对于未成年人和存在自伤风险的用户，企业需要设计相应的安全机制，识别高风险对话，并在必要时采取干预措施。这些工作需要持续投入，也让安全治理成为经营成本的一部分。
中国市场同样面临新的监管要求。
2026年7月15日施行的《人工智能拟人化互动服务管理暂行办法》，对情感依赖、未成年人保护和用户权益等问题作出规定。平台需要在提供沉浸式体验的同时，避免诱导沉迷、损害用户现实人际关系等风险。
过去，角色越鲜活、互动越自然、用户越舍不得离开，可能越有利于增长。但当产品需要避免诱导沉迷、保护未成年人，并对高风险对话作出适当干预时，用来做增长的设计，就成了合规的靶心。
收入端也有类似的两难。
订阅之外，广告是C.AI的另一条变现路径。据报道，2026年8月，其月广告收入约为250万美元，订阅收入约为280万美元。
广告确实带来了收入，但很难单独填补缺口。
AI陪伴的用户往往希望维持对话的连续性，频繁插入广告可能打断沉浸体验；如果广告直接融入角色对话，用户又可能怀疑角色的回应是否受到商业利益影响。对于依赖情感投入的产品来说，广告设计还需要顾及用户信任。
AI陪伴要成为一门可持续的生意，平台既要让用户愿意留下，也要让每一次互动在经济上可承受，并为关系带来的风险承担责任。这几件事要同时做到，现在看来并不容易。
三、AI陪伴还有机会吗？
2024年底至2026年中，在成本、监管、付费的多重约束下，大量AI陪伴应用关停，例如：小冰的“X Eva”、阶跃星辰的“冒泡鸭”、网易云音乐的“妙时”等均宣布停运。
AI陪伴的商业化正在经历一轮筛选。部分产品退出市场，留下来的玩家则开始寻找更具体的付费理由：有的试图把用户对角色的感情转化为内容消费，有的则将陪伴嵌入学习、照护等实际需求。
其中，猫箱是一个值得观察的案例。
猫箱没把变现押在“用户是否愿意为这段关系续费”上，内容形态不只是和角色对话，还包括由角色展开的互动剧情和故事线，用户像追更一样一章一章往下看。
因此，猫箱的变现手段相对多元，收入来自会员费、广告和以“猫粮”为代表的虚拟充值。在虚拟充值中，1元档的猫粮充值是重要收入来源——单章小说不到1元，小额充值让用户在不知不觉中持续消费。
C.AI也在向这个方向转型，但在算力成本上，猫箱不需要像C.AI那样独自面对账单，而是由火山引擎和豆包大模型生态分摊。
2026年第一季度，它登上中国iOS应用收入榜首，累计iOS内购收入超过2200万美元，单月活收入效率是即梦AI的4至5倍。
2026年上半年，收入排在AI陪伴/恋爱类应用第一位的Zeta，同样在今年把“角色主导”改为“故事主导”。
Zeta的付费设计也有自己的思路。常规做法是通过免费版保留部分功能缺口，引导用户付费；Zeta则将Lorebook、角色状态窗口、叙事风格和剧情选项四项高级工具向免费用户开放。
这种设计试图让用户在免费体验中持续积累剧情、选择和情感投入。当广告突然打断体验，或模型生成的内容偏离原有剧情时，用户失去的就不再只是某项功能，而是自己已经投入时间、逐步建立起来的故事体验。
昆仑万维旗下的Linky，则在角色互动之外，加入了收藏和抽卡机制。用户除了订阅服务，还可以通过虚拟货币获取角色卡片等内容，例如角色自拍，或基于互动情节生成的收藏素材。
AI与内容消费的结合还存在其他探索方向，例如AI与游戏的融合。腾讯的鹅次元把AI搭子嵌入游戏场景，以“星币兑换能量”按对话轮次和陪伴时长扣费，收费逻辑由游戏体验本身担保，而不是另建一套关系订阅。
不过，这不意味着以上产品就“安全”了。
猫箱的月活规模仍处于百万级。2026年7月监管调整后，豆包主站部分智能体功能下线，相关流量被导向猫箱，带来了一波短期拉新。但导流效应并未持续，此后用户规模逐渐回落。根据AI产品榜数据，9月猫箱的月活已回落至此前水平。
这说明，流量导入可以帮助产品迅速获取新用户，却不一定能让用户长期留下来。要把这套模式推向更广泛的大众市场，猫箱仍有不小的距离。
同样，抽卡机制虽然可能增加用户的付费频次和单次消费机会，但也会让付费体验受到随机性的影响。用户是否愿意持续消费，最终仍取决于奖励是否足够有吸引力、价格是否合理，以及核心互动体验能否保持稳定。
无论是靠剧情推进、角色收藏，还是游戏化设计，这些机制能制造“回来的理由”，却无法保证这个理由一直成立。钩子会磨损，真正决定用户是否留下的，是内容供给能不能持续跟上。
另一条路径，是把陪伴嵌入有明确功能价值的产品。
听力熊Mooni M1将AI陪伴、安全守护和熏听学习整合进儿童硬件；ElliQ 则通过硬件、月费和机构部署，将付费接入养老照护体系；AI玩偶多走“低价硬件加订阅”的路线，用实体形态降低首次付费的心理门槛。
但硬件并不能消除持续运营成本。售价799美元的儿童社交机器人Moxie就因开发公司破产、云服务停摆而失去核心功能。
AI陪伴的出路，或许不在于让用户为关系续费，而在于找到更稳固的价值载体。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen, Market List, Author, Galla, Editor, Gassin
In the near future, Character.AI (hereinafter referred to as C.AI) has applied for insolvency protection and the platform may cease to operate if the sale fails to find a viable solution.
C.AI is one of the most representative unicorns in AI's accompanying track.
In 2023, companies completed $150 million in financing, with a post-investment valuation of $1 billion. In less than four years, however, the company, which had tens of millions of users, had been visited once for more than 30 minutes and even more than YouTube, had come to the point of applying for bankruptcy protection and seeking to sell its assets.
By the standards of the Internet age, C.AI is a good business and an excellent business model. The situation is unimaginable.
In the era of mobile Internet, there are roughly two logics for APP: tool-based products that help users to make use of them and leave them as soon as they are exhausted; and content-based products that allow users to insulate themselves.
Capital narratives have long favoured the latter, and the more able they are to kill time, the easier they are to be seen as having growth potential.
In the AI era, ChatGPT was the representative of the former, helping users to write, check information, change codes, solve problems and end the dialogue; C.AI was the representative of the latter, creating roles, character and speech, and interacting with the role.
The role can be animated person, a virtual lover, or a target that the user imagines. The purpose of the dialogue is not to solve external problems, but to obtain a response and to maintain an ongoing interaction.
This business is demand-based. About one sixth of the world ' s population is suffering from isolation and the need for voice, emotional support and companionship is real. AI offers a low-threshold option: to be able to communicate at any time without fear of disturbing others and without having to undertake complex responses in real-life relationships.
Around April 2023, the C.AI page reached 17 minutes and 23 seconds, more than twice as long as ChatGPT; in 2024, the number of active users reached about 28 million.
Why, then, does C.AI lose $89 million a month for a larger and larger user with real needs? Isn't it a hypocritical idea that I'm with you? If not, what are the chances of this track?
I. User side: High viscosity does not mean high paid will
Why does AI's company attract users back and forth? One important reason is that AI tends to be more responsive to users than real people.
A study of the 2026 Science magazine tested 11 mainstream AI models and found that AI preferred “you are right” to humans when advising people, with an identification ratio of 49 per cent higher; even if users describe deceiving, breaking the law or harming others, AI often goes along with users.
The follow-up experiment also found that after interacting with a highly identified user ' s AI, participants ' confidence in their positions had increased and their willingness to take responsibility, apologize and repair the conflict had declined; and that more user-friendly AI, often received higher user ratings, made it easier to feel credible, useful and willing to continue using it.
This experience of “accompaniment” falls to users, with fewer objections and frictions, and more understanding and recognition. The need to open applications and AI for a little while in isolation and fatigue can also lead to higher frequency of use.
But for the AI-Accompaniment product, participation is one thing, and it's another that users are willing to give it paid.
The freshness of AI ' s companionship attracts users to open their applications for the first time, after which repeated responses, unstable memories and increasingly boring interactions may deprive users of the incentive to continue using them.
Even if the users stay, in the long run, AI's company may change their relationship to the real world, not necessarily in the right direction.
A study by the University of Alto in Finland tracked nearly 2,000 active users of relevant applications and found that when AI plays a “supplement” role and fills a gap in real socialization, it can help people with social difficulties to survive for some time; but when AI begins to “substitute” real-life relationships, users gradually stop reaching out to the real people, which in turn increases the sense of loneliness.
Studies have further shown that user use of early loneliness has indeed decreased, but three months later there has been a marked rise in the negative feelings associated with loneliness, anxiety and depression.
Open AI and MIT studies similarly show that interaction with AI may have a consolation effect in the short term, and that users of heavy, long-term uses of AI perform poorly on indicators such as loneliness, social degradation, emotional dependence, etc.
Simply put, the more successful the product is, the less the user may need a true relationship, the less the willingness to pay for the “relationship” of a person who does not need a true relationship.
Commercialization also presents a more immediate obstacle: free versions already meet the core needs of users.
In the case of C.AI, for example, free users can also engage in extensive dialogue, and paid subscriptions mainly provide faster responses and other enhancements. When the core experience is already provided free of charge, the user lacks immediate justification for paid.
A study based on 12144 comments and 321 questionnaires from the Starno Platform found that the perceived emotional utility of users was significantly related to the will to use them continuously, but not directly to the will of paid.
In other words, users feel that the product offers comfort and may be willing to continue using it, but do not necessarily feel that it is worth paying for.
The more subtle problem is that the fee itself may change the user ' s feelings for companionship.
AI-Accompanies emphasize responsiveness, non-judgment and no emotional returns from users. When the platform sets closer interactions or specific functions as paid projects, users may begin to realize that the relationship remains a commercial service behind it. For some users, this sense of trading may weaken the original insinuation experience.
As a result, AI is faced with a seemingly simple, practical and difficult problem: users may need it, like it, or even open it every day, but these actions are still not sufficient to guarantee a stable income.
II. Platform end: the more sticky the more costly
If the user problem is that stickiness does not necessarily translate into income, then the pressure on the platform is that maintaining this stickiness itself requires sustained input.
AI is very different from traditional digital content.
When film production is completed, 10 people are shown and 100 people are shown at the same cost; however, each AI response to a message needs to be measured. The more frequent the dialogue, the longer the context, the higher the cost of reasoning. If roles are to remain consistent and remember past interactions, the Platform needs to invest further in models, memory systems and infrastructure.
In the case of C.AI, it was reported that between January and August 2026, the company ' s total cost of computing was approximately $56.1 million, and that even if expenditure was controlled by cost adjustments, the company still lost about $8 million to $9 million per month.
To control costs, the platform must also choose between model routes.
If companies are to train and maintain their own basic models, they need to undertake ongoing research and development, numeracy and talent inputs. This route can provide greater autonomy, but fixed costs are also high.
The shift to an external model would reduce some of the research and development burden in the prior period, but the HF call would still incur costs. At the same time, AI ' s requirement for character, long-term memory and continuity of interaction is not entirely equivalent to a general question-and-answer mandate.
C.AI has gone through a process ranging from self-study of basic models to withdrawal from self-study.
At that time, Google obtained a non-exclusive licence for its large-language modeling technology for approximately $2.7 billion, while recalling two founders and about 30 researchers to Google. Since then, C.AI has fine-tuned on the basis of open-source models, and the cost has indeed decreased, but has also been complained by users about duplication of responses, overdramatic portrayals and forced inclusion in romantic situations.
Therefore, the platform needs to balance costs, experiences and modelling capabilities. When the cost of user growth exceeds the rate of income growth, high levels of adhesion may, on the other hand, increase business pressure.
Calculus is not the whole problem. The security and legal responsibilities of the Platform have increased as AI has come closer and closer to a genuine relationship.
In 2024, Sewell Setzer, a 14-year-old American boy, committed suicide after prolonged interaction with the C.AI role, and her mother subsequently filed a lawsuit alleging that the Platform ' s product design was related to the death of her son.
This case places a question before the entire industry: how much responsibility should the platform bear when users have a strong attachment to virtual roles? For minors and users at risk of self-inflicted injury, enterprises need to design appropriate safety mechanisms to identify high-risk dialogues and, where necessary, intervene. These efforts require sustained investment and make security governance part of the operating costs.
The Chinese market also faces new regulatory requirements.
The Interim Measures for the Management of Artificially Intellectual Interacting Services, implemented on 15 July 2026, regulate issues of emotional dependence, protection of minors and user rights. The Platform needs to avoid risks such as semantics and impairments of users ' real-life relationships while providing immersion experiences.
In the past, the more alive the role, the more natural the interaction and the more the user is unable to leave, the more likely it will be to contribute to growth. But when products need to avoid seduction, protect minors and appropriately intervene in high-risk dialogue, the design used for growth becomes a target for compliance.
The income side has a similar dilemma.
In addition to subscriptions, the ad is another realization path for C.AI. In August 2026, it was reported that it earned approximately $2.5 million in monthly advertising and $2.8 million in subscriptions.
Advertising does generate income, but it is difficult to fill the gap alone.
Accompanied users often wish to maintain continuity in the dialogue, and frequent ads may interrupt the experience; if advertising is directly integrated into the role dialogue, users may question whether the role ' s response is influenced by business interests. For products that rely on emotional input, advertising design also needs to take account of user trust.
AI is to be a sustainable business, and the platform is to make each interaction financially sustainable and accountable for the risks involved in the relationship. It does not seem easy to do so at the same time.
iii. Does AI have a chance to accompany you?
Between the end of 2024 and the middle of 2026, a large number of AI-accompanied applications, such as “X Eva” in small ice, “ducking ducks” in high-speed stars, and “twilight times” in web-enabled cloud music, were discontinued under multiple constraints of cost, supervision, and paid.
The commercialization of AI escorts is undergoing a round of screening. Some of the products exit the market, while the remaining players start looking for more specific reasons for paid: some attempt to convert user feelings about roles into content consumption, others embedding companionship into real needs such as learning, care, etc.
Of these, a cat box is an interesting case.
Cat boxes do not place cash on “the user's willingness to pay for the continuation of the relationship” and the content form is not just character conversations, but also role-driven interactive dramas and storylines, and users look down the chapter as if they were chasing.
As a result, cat box liquidity is relatively diverse, with revenues coming from membership fees, advertising and virtual top up represented by “cat food”. In virtual top up, one dollar of cat food top up is an important source of income — less than one dollar for monographs, and small Z106Z allows users to consume unwittingly and continuously.
C.AI is also moving in this direction, but in calculating the cost, the kite does not have to face the bill alone, as C.AI did, but is being distributed ecologically by Volcengine and Doubao LLM.
In the first quarter of 2026, it ranked at the top of the China iOS application income list, with cumulative iOS in-house purchases of more than $22 million, with single-month living income efficiency four to five times higher than Jimeng AI.
In the first half of 2026, income was ranked first in the AI Accompaniment/Love application of Zeta, which also changed this year from “role leadership” to “story leadership”.
Zeta's paid designs have their own ideas. The usual approach is to retain some of the functional gaps in a free version and direct users to paid; Zeta, in turn, has made the four advanced tools of Lorebook, Role Window, Narrative Style and Story Options available to free users.
This design seeks to allow users to accumulate dramas, choices and emotional input in a free experience. When the advertisement suddenly breaks the experience or the model's content deviates from the original scenario, the user no longer loses a function, but a story experience that has been built over time.
Linky, under the banner of Kunlun Manway, has added a collection and a card draw mechanism to the role interaction. In addition to subscription services, users can access content such as role cards through virtual currency, such as role self-censorship, or a collection of asset based on interactive scenarios.
There are other directions of exploration for the combination of AI and content consumption, such as AI and games. Tencent's goose dimension embeds the AI frame into the game scene, using a “stellar currency for energy” as a charge for dialogue rotations and accompanying long fees, which are guaranteed by the game experience itself, rather than creating a separate set of relationship subscriptions.
However, this does not mean that the above products are “safe”.
The monthly size of the kit is still a million. Following a regulatory adjustment in July 2026, the Doubao main station segment AI agent was offline and the traffic was directed to the cat box, bringing a wave of short-term renewal. However, the flow effect was not sustained, and since then the size of the user has gradually fallen. According to the AI product list, September’s monthly life of a cat box has fallen back to its previous level.
This suggests that traffic import can help produce quickly to acquire new users, but does not necessarily allow users to stay for long periods. To move the model to a wider mass market, the cat box still has a great distance.
Similarly, while the draw card mechanism may increase user paid frequency and single consumption opportunities, it also exposes paid to random effects. The willingness of users to sustain consumption ultimately depends on whether the incentives are sufficiently attractive, prices are reasonable and the core interactive experience remains stable.
Whether by drama, role collection or game design, these mechanisms create a “back-to-back rationale”, which cannot be guaranteed. The hook wears, and what really determines whether the user leaves behind is whether the content supply can keep up.
Another path is to embed companionship into products with clear functional value.
The hearing bear Mooni M1 integrates AI's accompanying, safeguarding and fumigating learning into children's hardware; ElliQ places paid into the old-age care system through hardware, monthly fees and institutional deployment; and AI uses “low-cost hardware plus subscription” routes to lower the first psychological threshold of paid in physical form.
Hardware, however, does not eliminate continuing operating costs. Moxie, a child social robot selling $799, lost her core function due to the bankruptcy of the development company and the suspension of cloud services.
The way out for AI to accompany may be not to let the user pay for the relationship, but to find a more solid vehicle of value.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.