China's AI Shakeout: 38 Tools Shut Down as Models Absorb Apps
ByteDance and Tencent's agent platforms join a sweeping consolidation that exposes the structural limits of China's first-generation AI app economy.
ByteDance's AI agent development platform Coze Enterprise Edition will stop accepting new customers on November 30, 2026, with all existing subscriptions mandated to expire by December 31, 2027—a quiet but telling retreat from a market segment Coze once led. The announcement, paired with Tencent's separate confirmation that its Yuanqi agent platform will go dark on November 9, 2026, crystallizes a structural reckoning reshaping China's AI application landscape.
According to an incomplete tally by Tech Planet, at least 38 AI products across China's technology sector have been shut down or absorbed into parent platforms, with 25 of those originating from major internet conglomerates including Alibaba, ByteDance, Tencent, and Baidu. The pace of closures has sharply accelerated since mid-2026, compressing what once looked like a broad-front product expansion into a narrower, more defensible set of core AI assets.
Coze and Yuanqi Closures Signal Model Intelligence Obsoletes the Middleware Layer
The near-simultaneous shuttering of two of China's most prominent agent-building platforms is not coincidental—it reflects a convergent logic rooted in model capability curves rather than commercial mismanagement.
Coze Enterprise launched in the second half of 2025 and operated for just over a year before entering its wind-down. Yuanqi, released in May 2024, ran for roughly two and a half years. A former ByteDance Coze employee, posting publicly, framed the closure succinctly: the product was not displaced by AI, but by the model's own growing capacity for autonomous planning. As frontier models internalize multi-step reasoning and tool orchestration, the low-code workflow layer that platforms like Coze and Yuanqi occupied becomes redundant.
The commercial context amplifies the pressure. IDC's 2025 China Intelligent Agent Development Platform Market Share Report ranked Coze first in China's public cloud agent development market with a 19.3% share—yet the entire market generated only RMB 250 million (approximately US$34.7 million) in that year. Market leadership in a structurally constrained segment offers limited insulation against obsolescence.
Big Tech Consolidates Around Three-Layer Architectures, Retiring Dozens of Peripheral Products
The restructuring moves among China's large platforms follow a recognizable template: collapse fragmented experimental product lines into a tiered architecture of consumer app, enterprise productivity suite, and cloud infrastructure.
ByteDance's strategic geometry is now explicit. Consumer AI is anchored in Doubao; enterprise productivity runs through Feishu and Doubao Work; infrastructure and enterprise delivery are handled by Volcano Engine. The products that did not fit—Volcengine Writing, MarsCode (now TRAE), the AI camera app Xinghua, and Coze Enterprise—have been absorbed or retired across a 24-month period ending in late 2026. On July 30, ByteDance internally announced the merger of Feishu and Doubao product teams. By August 24, the TRAE and Coze teams had been fully folded into Doubao.
Tencent's consolidation has been equally sweeping. The company's near-decade-old translation product Fanyi Jun was decommissioned on March 13, 2025, with capabilities migrated to Yuanbao. Digital avatar tool Zhiying cleared all user data by June 30, 2025. QClaw which attracted over 100,000 pre-registration sign-ups upon its March 2026 launch, received a shutdown notice just six months later, with services terminating December 24, 2026. Its business unit was redirected into WorkBuddy (WorkBuddy) in July 2026.
Alibaba shuttered its decade-old AI design product Luban on June 30, 2025. Tongyi Tingwu lost its standalone app identity and was merged into Qianwen. Three separate agent product lines—QoderWork, MuleRun, and Wukong—were consolidated into Qianwen Office on August 3, 2026.
Baidu retired its ad-free AI search product Tizi AI on March 4, 2026, just six months after launch, folding it into Wenxin. Wenxin Yige and Wenxin Kuaima were similarly absorbed into Wenxin Yiyan and Baidu Dazi, respectively. Baidu's superintelligent agent app Xinxiang, unveiled at a developer conference, disappeared from public visibility within three months.
Startups Face Harder Arithmetic: No Internal Safety Net Absorbs Failed Bets
Where large platforms can redeploy teams, users, and intellectual property across internal businesses, venture-backed AI companies operating failed product lines face binary outcomes.
Moonshot AI tested overseas role-play product Ohai and audio-video generation tool Noisee in the first half of 2024, halting iteration by September of that year and redirecting all resources to Kimi. The pivot did not fully resolve capacity constraints: following the launch of Kimi K3 in 2026, user request volumes reportedly approached compute limits.
StepFun's character-agent product Maopaoya achieved over six million downloads across platforms after its April 2024 launch, but was folded into the company's core model development operation and ceased service on November 5, 2025, as the company shifted investment priority toward foundational model research.
01.AI announced on August 3, 2026 that it would close its public-facing large model platform, terminating external API services by September 3. The move marks a full pivot to B2B enterprise AI solutions, abandoning consumer and developer-facing products including virtual companion MonalandAI, productivity tool Wanzhi, AI search app Beago, social app Shado, and overseas social product Bingo.AI.
MiniMax (MiniMax) rebranded its Hailu AI under the MiniMax name, redirecting the original brand toward video. Its virtual character app Glow (Glow) has endured a turbulent regulatory history: removed in March 2023 for compliance reasons, re-launched as domestic product Xingye and overseas product Talkie, with Talkie itself suspended by Apple for two months in 2024. Zhipu AI's video generation product Qingying was absorbed into Zhipu Qingyan and its domestic companion product AiU (AiU) was shut down.
Three Structural Forces Are Rewriting the Viability Equation for Standalone AI Tools
The consolidation wave is not random. Three distinct mechanisms are driving product exits, and each suggests different risk exposures for companies still operating independent AI applications.
Model capability absorption is the most pervasive. Functions that once justified standalone products—translation, writing assistance, image generation, information retrieval—are increasingly native capabilities of general-purpose models. The manual workflow configuration that platforms like Coze required users to perform is being superseded by models that can autonomously parse a task objective, decompose it into steps, invoke the relevant tools, and return a complete output. Single-function AI tools lack the moat to survive this transition.
Regulatory boundary-setting has sharply compressed the AI companion and agent customization segment. Five Chinese government agencies issued the Interim Measures for the Administration of AI Anthropomorphic Interactive Services in April 2026, effective July 15, explicitly designating "emotional dependency" as a regulatory red line. Within days, Doubao, Qianwen, and Yuanbao simultaneously removed user-created agent features, affecting over eight million characters. The compliance cost-to-revenue ratio proved unworkable for platforms whose differentiation relied on customizable persona interaction.
Unit economics failure has eliminated a third category. AI application cost structures tie inference spend to usage volume, inverting the typical SaaS model: high engagement drives losses rather than margins. Character.AI, with 12 million active users globally, reportedly sustains monthly losses approaching US$10 million. With user payment conversion rates in China's AI app market remaining low relative to compute cost escalation, the business model underpinning many consumer AI tools is fundamentally unsound.
Supply-side dynamics compound the pressure. As of August 31, 2026, cumulative registered generative AI services in China had reached 1,112, up from 302 at end-2024, with an additional 731 generative AI applications or functions completing registration. As undifferentiated supply multiplies, the scarcity premium that once attached to any AI-powered feature has evaporated.
Impact Assessment: What Survives the Consolidation
The pattern emerging from this consolidation suggests that viable standalone AI products in 2026 require three simultaneous conditions: anchoring in a high-frequency use case, a paying user base with demonstrated willingness to absorb subscription costs, and the ability to deliver verifiable business outcomes rather than capability demonstrations.
Products that satisfied only one or two of these conditions—even those with strong initial user acquisition metrics like QClaw's 100,000 pre-registrations—have proven insufficient to withstand the combined pressure of model capability advancement, tightening regulation, and deteriorating unit economics.
For China's AI ecosystem, the first wave of experimentation is closing. The second wave will be defined not by breadth of product launches, but by depth of integration into workflows where switching costs are real, usage is habitual, and the value delivered is quantifiable. ByteDance, Tencent, Alibaba, and Baidu have each placed their bets on what that architecture looks like. The score will be settled in enterprise contracts and consumer retention data through 2027.
Related Coverage:
Unitree's Wang Xingxing Sees Physical AI Breakthrough Within Two Years
AI-Powered Games Flood China's Market, But Quality Gap Widens as Platforms Race for Scale