Original · Unique Research · 2026-01-19
Historical edition: This is the full textual English rendition of Unique Research’s January 19, 2026 article about December 2025. First-person commentary, forecasts, product capabilities, rankings, institutional descriptions, methodology and accuracy claims are the original author’s statements at that time, not independently verified current findings. The source’s definitions of unique visitors and subscription revenue (labelled ARR) are reproduced as its own methodology, not substituted with other industry definitions.
When AI Finally Learned to “Go to Work”: 13 Products on the Growth Ranking Are Turning Chat Boxes into Assembly Lines
Over the past year, we have witnessed a “Cambrian explosion” in the AI industry. Countless applications have sprung up, most wearing a similar face: an input box, a blinking cursor, and the greeting, “How can I help you?”
EXECUTIVE INSIGHT People are leaving “chatbots that speak in generalities” for tools willing to reach into workflows and take over the dirty, exhausting work. These products are in no hurry to prove how well they can talk. Instead, they repeatedly answer a more practical question: beyond chatting with me, can AI help me finish this job?
The ranking for the final month of 2025 became especially interesting. After deduplicating the AI Web growth ranking against the Top 10 monthly-active-user growth ranking, 13 applications remained, like a silent yet consistent vote. If you treat these 13 products as scattered points, they appear widely dispersed: no-code application creation, task-oriented agents, professional editors, enterprise collaboration, information organization, and real-time voice engines. But if you view them as a shared tide, a larger shift becomes visible: AI is moving from “conversation” to “execution,” from “generation” to “delivery,” and from a “moment of amazement” to “long-term compounding.”
13 SELECTED GROWTH PRODUCTS
The moment barriers collapse, “making things” begins to feel as natural as breathing
Many technological waves truly change the world not because they make professionals another 20% faster, but because they give nonprofessionals the ability to turn expression into a finished work for the first time. MeDo exemplifies this collapse of barriers: describe an idea in natural language, and it gives you not a concept image but the beginnings of a working application, which can even be published in an environment like a Marketplace, where people can access, try, and modify it. You suddenly realize that ideas once blocked by “I do not have development resources” can now be pulled into reality through a usable prototype.
OiiOii.ai goes one step further. It breaks the highly collaborative, high-barrier work of animation creation into an AI team you can direct: screenwriters, storyboard artists, character designers, scene designers, music creators, and more. You are not operating a tool so much as directing. All you need is a vague world or a one-sentence plot description, and it can advance an “idea” into a “watchable clip.” This may not constitute industrial-grade delivery, but it is enough to let first-time users taste a dangerously appealing realization: I can make something too.
Lingguang takes a gentler approach, placing itself at the moment when “inspiration has just appeared.” It does not wait until your task is fully defined, but catches you when your thinking is still unclear and you have only a passing spark. Through “Flash Apps” (闪应用, a translation of the source’s term), it reduces much of the recurring cost of content production: a structure, template, or mode of expression generated once can be reused the next time. It is like a container that gives unfinished ideas somewhere to land.
These three products look very different, but the same growth logic lies behind them: when the psychological cost of “making something” approaches zero, the speed of experimentation rises exponentially. Ideas no longer remain in the mind; they rapidly become tangible prototypes. Growth is often hidden in that moment of “making it for the first time”—an addictive sense of control.
From Answer to Deliver: agents stop chatting and take over the process
What truly gives the ranking a sense of the times, however, is another category of product: these applications almost all deemphasize “conversation” itself and replace the basic unit of interaction with the “task.”
AutoGLM, a task-oriented agent, is a straightforward signal. You do not chat with it; you give it a goal: summarize a video, generate a PPT, or organize a pile of material into a webpage. It tries to absorb all the intermediate steps, allowing you to step back from being a “prompt engineer” and become more like the person accepting delivery on a project—you define the expected outcome, and it completes the journey.
The change in Qwen is equally important. Instead of resembling the model-showcase window it was early on, it has been repositioned as an “AI workbench that can do the job directly.” Its entry points have been divided into writing, translation, PPT creation, document processing, audio and video understanding, image generation, coding, deep thinking, deep research, and more. This decomposition looks simple but is extremely effective: first-time users do not need to learn “how to ask.” They can enter directly through “what do I need to deliver?” A product that lowers this exploratory cost can win many incremental users who do not care about model parameters and care only about saving time.
Oreate AI turns this idea into a more explicit form: it packages capabilities such as Slides Agent and Deep Research into vertical agents, making the “deliverable” the default goal of interaction. Flowith also incorporates the dimension of time, emphasizing an Agentic Workspace and a Knowledge Garden; many complex tasks cannot be completed in three sentences. They require AI to possess “patience” and “continuity,” moving work forward and refining it repeatedly in the same space, and transforming one-off questions and answers into a project site where work can continue.
When you view these products together, you discover that they are doing the same thing: organizing AI into a system that can work, rather than an interface that can talk. Users never came to AI merely to chat—they came to get things done.
Generating is easy; editing is hard: the “second brains” of professional tools are taking shape
If the previous stage of AI offered the “thrill of generation,” competition in this stage is quietly returning to a tougher question: what happens after generation?
Kira’s ambition lies here. Unlike many image tools that treat AI as a standalone button, it moves the entire workstation of a professional editor into the browser: layers, masks, lassos, brushes, blend modes, and more, all as familiar as Photoshop; AI behaves more like an embedded capability here: expanding the canvas, erasing elements, lossless upscaling, and context-aware generation. You do not ask AI to “give me an image”; you ask it to “do one thing” at this specific location in the image. More important, generated content becomes a layer that remains editable, eliminating the break in continuity. For people who genuinely design, this continuity is more valuable than any “generation speed.”
Baidu Comate is doing something similar in the world of code. It is not satisfied with completing a few lines. It tries to understand the context of entire files and entire projects and perform more “engineering-oriented” collaboration inside an IDE (integrated development environment): implementing changes from requirements, refactoring, adding unit tests, adding comments, modifying similar code in batches, and even decomposing and integrating work through multi-agent collaboration. It advances AI from a “code search engine” to an “engineering-level colleague,” while bringing practical concerns such as security compliance and on-premises deployment to the foreground. Those concerns are precisely the dividing line between whether enterprises are willing and able to deploy it at scale.
Inworld AI represents a more foundational direction. It resembles a runtime engine for real-time AI applications, emphasizing Realtime TTS (real-time text-to-speech), low latency, emotional expression, scalability, and integration across the runtime and orchestration layers. When you are building not something that “reads a script” but an AI character capable of sustained interaction, the foundation cannot be a single call to a large model; it must be an engineered combination of an entire chain. Inworld does not compete for center stage, but it may stand behind many stages.
Ruliu applies this “reliability first” philosophy to organizational collaboration. It does not try to create short-lived amazement with dazzling features. Instead, it consolidates access points, connects processes, and stabilizes governance: compatibility with domestic technology stacks, on-premises deployment, security audits, China’s multi-level cybersecurity protection framework, and more. These terms may not sound exciting, but they are the reasons medium-sized and large organizations are genuinely willing to pay and keep using a product over time.
Turning information from Flow into Stock: tools that fight information entropy often look the quietest
The growth ranking also reveals a quieter but potentially more far-reaching need: people are no longer satisfied with “acquiring more information.” They increasingly need to “keep information from going to waste.”
Twillot’s entry point looks small: anonymously viewing posts, offering a clean interface, and downloading media. Its real value, however, lies in turning Twitter’s “scroll and leave” information flow into a personal archive that can be synchronized, indexed, and exported—even to Markdown, CSV, PDF, or Obsidian for continued use. It is a reminder that social platforms are increasingly adept at manufacturing the “present,” yet do not care whether you can own the “past.” When platforms refuse responsibility for your long-term value, tools emerge to return your data to you.
YouMind addresses the same broken link: the gap between input and output. It lets you casually collect webpages, PDF files, videos, podcasts, and ideas, but goes beyond summarizing. It generates insights around your highlights and annotations, then moves those insights into editable documents, naturally leading you into “revising, expanding, and rewriting.” It does not learn for you; it is more like a force compelling you to turn learning into expression—and expression is when information truly begins to compound.
The “execution flywheel”: an AI product’s value is determined not by intelligence, but by its delivery radius
If you separate these 13 products into their respective categories, the picture becomes fragmented. Place them back on the same evolutionary path, however, and a more general framework emerges. I prefer to call it the “execution flywheel,” because it describes not a single feature but a structure capable of generating sustained growth.
The first turn of the flywheel is a low-friction entry point: natural language, outcome orientation, task templates, and no-code generation make a first use require almost no learning. Growth for MeDo, Qwen, and AutoGLM begins here—bring users in and let them accomplish something once.
The second turn is persistent context: complex tasks take time and repetition, and they need a workspace that does not reset to zero simply because a “conversation ends.” Flowith makes this most explicit by allowing AI to “patiently” accompany you as you gradually finish the same project.
The third turn is controllable editing: generation is only the starting point; real value often emerges through revision, refinement, and engineering detail. Products such as Kira, Baidu Comate, and Ruliu embody long-term thinking because they put AI into the “least human-friendly parts” while leaving ultimate control with people.
The fourth turn is asset accumulation: every use can leave behind a reusable result, creating a positive cycle. The value of Twillot and YouMind lies in turning information and time back into your assets, rather than leaving you as a passerby on a platform.
AI will become increasingly “boring,” but also increasingly useful
Looking back at these 13 names, the most exciting point is not whose features are more elaborate, but the future they collectively anticipate: AI is retreating from center stage to backstage, changing from a soloist into an executor and from an “amazing tool” into “everyday infrastructure.” It will look less and less like a black box to be admired, and more like a button at your workstation, a browser plug-in, or a colleague in the editor’s sidebar.
The real questions therefore become sharper. Once AI finally learns to work, which jobs are we willing to hand over to it? Which steps must we complete ourselves? And what content is worth accumulating as an “asset,” rather than consuming as an “information flow”?
This may be what the growth ranking most wants to tell us: the AI era no longer proves itself by “being able to talk”; it remains by “being able to deliver.” You no longer need to seek a smarter AI. You need to find one that makes you more like a professional, more like a creator, and more like yourself.
01. About Us
Unique Research — An Authoritative Institution for Global AI Market Insights
Our Influence and Recognition
Unique Research’s data and insights have earned broad trust and citations from leading investment institutions, academic organizations, and media outlets in China and abroad, becoming an important benchmark for market analysis, investment decisions, and trend assessment.
Our partners and organizations that cite us include, but are not limited to:
• Securities firms and investment banks: Sinolink Securities, China Securities (中信建投), Huayuan Securities, Haitong Securities, and others
• Authoritative media: China Securities Journal, TMTPost, Jiemian News, and others
We firmly believe that only data able to withstand verification by core industry participants has genuine long-term value. We believe in the power of open source and openness, and Unique Research therefore makes the following three commitments:
1. The AI Top 100 ranking will remain permanently open-source and freely available, with its complete methodology and raw data supporting full reproduction.
2. Data in the AI Top 100 ranking is strictly limited to independent third-party monitoring platforms; self-reported company data is not accepted as an evaluation basis.
3. The ranking system of the AI Top 100 remains absolutely neutral and accepts no commercial partnership or payment that interferes with ranking results.
How Can You Use the AI Top 100?
• If you are an investor/analyst: this ranking provides a clean data benchmark that can withstand reverse-engineering verification. It is a reliable foundation for building investment models, discovering under-the-radar projects, and performing cross-validation.
• If you are an AI entrepreneur/product manager: you can not only identify competitors precisely, but also reproduce our methodology in depth to analyze your own product’s market performance and guide optimization and iteration.
• If you are a researcher/student: its fully open-source nature makes this report an excellent free resource library for academic research and market analysis. We encourage innovative research based on this data.
For the complete data behind the charts and rankings shown in the AI Top 100, please visit 100aiapps.cn; all visible data may be copied and is freely available for noncommercial uses such as personal learning and research.
02. Research Statement
2.1 Data Description
The AI-product revenue data contained in this report, article, ranking, or chart is based on real-time tracking of referral traffic from official AI-product websites to payment gateways such as Stripe; combined with time-series analysis, it examines historical visits, fluctuations in monthly active users, and changes in pricing strategies to construct a dynamic revenue-estimation model. Sampling validation shows that the mean absolute error rate between estimates and actual data remains within ±10%, meeting standards for industry-grade research accuracy.
The AI-product traffic data contained in this report, article, ranking, or chart is based on the integration of multiple data sources and processing by Unique Research’s intelligent algorithms. Monitoring is strictly limited to users’ direct access through official applications and websites. Specifically, the data covers only direct user access to AI products through websites and native applications; it excludes browser plug-ins, desktop clients, mini-program ecosystems, embedded services within Discord, open-source local deployments, API calls, and other scenarios.
±10% PREDICTION ACCURACY MARGIN
2.2 Definition of Concepts
• Geographic Dimension
• Overseas AI products: founded by entrepreneurs or teams who are not ethnically Chinese and aimed primarily at the global market.
• Domestic AI products: founded by Chinese-national entrepreneurs or teams and aimed primarily at China’s domestic market.
• China-origin products going global: founded by ethnic-Chinese entrepreneurs or teams but targeting primarily overseas markets.
• Functional Dimension
• AI-native applications: products designed from the outset with AI technology deeply integrated, whose core value depends entirely on AI.
• AI-enabled applications: products that enhance existing functions by integrating AI technology into their original business logic.
• AI-ecosystem applications: platforms that provide support and connections for AI researchers, developers, service providers, and users.
2.3 Metric Description
• WEB Data Metrics
• Visits: the total number of all visits to a website. Renewed activity after an interval of more than 30 minutes is counted as a new visit.
• Unique visitors: the number of unique IP addresses that visit a target website during a specific period.
• Subscription revenue (ARR): all revenue a website earns from subscription services, excluding advertising and one-time revenue.
• APP Data Metrics
• Downloads: the number of times an App is downloaded by users.
• Active users: the number of unique users who perform at least one activity during a specific period.
• In-app purchase revenue (IAP): revenue generated by users purchasing virtual goods, premium features, and similar items.
2.4 Free-Use Statement
This report is published by Unique Research and is copyrighted by Unique Research. Any Chinese-language republication or quotation must identify the report as its source. When citing data, please state: [Data source: Unique Research]. This report is an independent original analysis, represents no company’s position, and does not constitute investment advice.