The Ones Paying for the World's First AI Agent Phone Might Not Be the Users

Author｜Yixiu

Editor｜Xiaobai

Cover Image｜Generated by AI

Produced by｜Qiangdiao Next

On the evening of July 13, Jieyue Xingchen released its AI terminal brand STEPX, the agent-native operating system Step AOS, and the personal agent Jieyue Amoo, with its first phone, STEPX Neo, also making its debut.

Four days later, at the World Artificial Intelligence Conference, agent phones will face their first direct competition.

Nubia will showcase a mass-produced flagship model that it claims to be the 'world's first AI agent phone,' widely seen as the second-generation Doubao phone. Honor also confirmed that its Robot Phone will be unveiled at the WAIC special forum on July 18 and will be available for on-site experience.

Together with Jieyue Xingchen's STEPX Neo, three models, all claiming to be the 'world's first,' will be displayed side by side. Jieyue emphasizes its native agent operating system, Nubia highlights mass production and cross-application execution, while Honor combines agents with a four-degree-of-freedom mechanical gimbal.

The fact that three 'world's firsts' have emerged in such a short time indicates that there is no universally recognized standard in this sector. Anyone can redefine 'first,' precisely because no one has truly succeeded yet.

01. AI Native Apps Are Booming, But Phone Manufacturers' AI Isn't Being Utilized

QuestMobile's '2026 Mid-Year AI Application Market Development Insights Report,' released yesterday, explains the underlying logic behind these manufacturers' Layout (strategic moves) into agent phones.

As of June, domestic AI native apps had 499 million monthly active users, an 85.4% year-on-year increase, with an average monthly usage time of 183 minutes per user, a 40% increase, crossing the threshold from 'novelty' to daily necessity. However, the same report shows that the largest volume actually comes from system assistants pre-installed by phone manufacturers, with 755 million monthly active users, 256 million more than AI native apps.

However, in terms of average monthly usage frequency, AI native apps are used 92.7 times per user, while system assistants are only used 51.4 times, nearly half as much, with an even larger 18.7-fold difference in usage time. Phone manufacturers have won in terms of volume through hardware pre-installation but have fallen behind in usage depth.

This contradiction is even more glaring at the brand level. QuestMobile data shows that OPPO's Xiao Bu ranks first with 165 million monthly active users, but the most frequent users are actually Xiaomi's Super Xiaoi (29 times per user per month). Apple's Siri is at the bottom, with only 3.8 times per user per month, less than one-twentieth of the average for AI native apps.

This exposes an awkward truth: phone manufacturers have pre-installed more and more AI features, but users have not changed their usage habits accordingly.

Summarization, translation, image editing, and search seem to cover everything, but they are just scattered tools within the system. Users still have to think of these AI features first and then actively call them up. No matter how powerful the model is, if the entry point is not established, capabilities are hard to translate into usage frequency.

Yin Qi, Chairman of Jieyue Xingchen, put it more bluntly, saying that agents residing in others' operating systems are forever just 'guests.' Only by building a new 'house' for agents can they become 'natives.'

But this seems more like a judgment aimed at the industry and capital markets rather than a conclusion that has already been verified.

The core change in the three prototype models is to elevate AI assistants to the next level, making AI not just a pre-installed feature but a system-level control center. However, the starting points and forms differ between phone manufacturers and model developers. Whether this path can succeed depends not on technology but more on commercial interests and user habits.

Although STEPX Neo has passed the national standard L3 test for 'Artificial Intelligence Terminal Intelligence Grading,' the certificate proves its capability ceiling but not whether users are willing to entrust it with complex cross-app operations—precisely the part that none of the three companies have yet answered.

02. The Routing Dispute Behind the Three 'Firsts'

The core issue lies in how agents can call upon third-party apps.

Jieyue follows a protocol-based approach, with its GUI-MCP protocol adopting a layered dual-stack architecture. Original screenshots remain local, with only semantic summaries transmitted to the cloud. The open-source GUI proprietary model has 4 billion parameters and can reportedly recognize and operate over 200 apps. Initial ecosystem partners include Alipay, Meituan, Didi, Gaode, JD.com, Baidu, WPS, CapCut, Ctrip, Tongcheng, Weibo, and others.

However, the market's most concerned apps—WeChat, Taobao, and Douyin—are not on the list.

Nubia takes the most aggressive approach. The first-generation Doubao phone (engineering model Nubia M153, released in December 2025, priced at 3499 yuan) relied on system-level 'INJECT_EVENTS' simulated touch permissions, with the wake word 'Doubao Doubao,' capable of responding to almost any user command. It quickly sold out after release, with second-hand prices once soaring above 10,000 yuan.

But the cost (cost) followed. This high-frequency automated clicking was soon identified by WeChat as abnormal operation, triggering 'environment abnormal' alerts and restricting logins for some accounts. WeChat, Taobao, and other platforms subsequently blocked relevant permissions.

Tencent later publicly stated its opposition to uploading user phone screen recordings to the cloud via automated plugins. According to reports, the second-generation model has negotiated permission openings for high-frequency scenarios like social, office, and travel with some leading internet companies, but the extent of these openings will only be clear on July 17.

Honor takes the most 'rule-abiding' approach. Its YOYO agent is based on a system-level MCP architecture, covering over 3,000 scenarios. It is secure but limited by whether third-party developers are willing to proactively adapt interfaces. If commands are too complex or span multiple unadapted apps, YOYO may ' Strike ' (go on strike).

This 'first judge capability boundaries, then issue commands' interaction logic partly explains why Honor's flagship Magic8 has not received strong market feedback. According to data disclosed by tech bloggers earlier, as of the 8th week of this year, Magic8 series sales stood at only 795,500 units, ranking last among flagship models from the TOP 6 phone brands.

03. Who Has the Authority to Click 'Confirm' for Users?

The fundamental difference between the three approaches lies in whether agents need App approval to act on behalf of users.

The logic of simulated clicking is that if users can operate something, agents should in principle (in principle) be able to as well, without waiting for any platform to open interfaces. This is the most aggressive and fastest path but also the most challenging. IDC predicts that 147 million AI phones will be shipped in China in 2026, accounting for over half, but more than 90% will only overlay AI assistance functions within the system without truly touching the underlying paradigm of human-machine interaction. The Doubao phone is one of the few products daring to take a step forward, and its cost (cost) is being the first to hit the boundaries of major apps.

The MCP protocol follows the opposite logic, letting Apps decide how far to open up. Jieyue's GUI-MCP and Honor's system-level MCP both fall under this path—safe and controllable but inherently slower. App developers are not obligated to adapt immediately, and the completeness of the user experience depends on ecosystem partners' willingness, not technical limits.

The problem is that both paths ultimately face the same dilemma: super platforms like WeChat, Taobao, and Douyin are almost impossible to relinquish control over their transaction and account systems. This is a question of who controls traffic allocation and advertising systems. While it involves security on the surface, it goes deeper to core business models.

Thus, what agent phones are truly competing for is the right to interpret and choose services on behalf of users.

Alibaba breaks down Taobao and Gaode into callable atomic skills, offering limited openness under its own rules. Tencent's restrained posture of 'light connection, heavy backend' essentially refuses to cede the power to interpret rules. None of the three internet giants are willing to budge first, making this the toughest nut to crack in this melee.

Without support from these major platforms, agent phones inevitably seem somewhat niche.

04. Phones Are Not Products; They Are Bargaining Chips

QuestMobile's half-year report already shows clear commercialization signals from AI native apps, not agent phones.

In June, Doubao's three application formats all saw steady activity growth, with the App version gaining 13.78 million new users compared to May. After launching a paid professional version on June 24, daily active users surged to over 178 million the next day, hitting a monthly peak. The paid model did not hurt the user base, with the 'inclusive + premium' approach already proven.

Meanwhile, Qianwen had 167 million monthly active users, a 5792.9% year-on-year surge, and DeepSeek had 130 million. Among the first tier formed by these three, the proportion of heavy users (10+ minutes per session) is rising. Doubao leads at 27.5%, DeepSeek at 30.0%, and Kimi at 26.1%, all up at least 2 percentage points year-on-year.

In contrast, the three agent phones are still stuck in the prototype verification stage.

As of July 14, STEPX Neo has not announced its specs, price, or release date. Jieyue Xingchen stated it will not monetize through hardware margins or pre-installed ads and has no rigid shipment targets. Nubia's mass-produced flagship has not announced pricing or release dates. Honor's Robot Phone has only confirmed a Q3 launch, likely around August. At this stage, agent phones seem more like declarations of industrial status than proven consumer electronics businesses.

What truly determines who will ultimately prevail lies in the industrial bindings beneath the surface.

In January, Jieyue Xingchen completed a B+ round financing of over 5 billion yuan, with Huaqin Technology among the industrial investors. In May, the company completed another round of nearly 2.5 billion USD, with Huaqin, Longcheer, OmniVision, ZTE, and other core consumer electronics supply chain companies joining, along with the 'Hong Kong version of Temasek,' Hong Kong Investment Management Corporation, entering the shareholder list. The company simultaneously completed share reform and removed its red-chip structure, with media judging this as a clear move toward a Hong Kong IPO.

It is noteworthy that Step series models from StepFun have been pre-installed in approximately 60% of leading mobile phone brands through collaborations with manufacturers such as OPPO, Honor, and ZTE, with installations surpassing 42 million units and serving nearly 20 million users daily. Therefore, the number of units sold for the STEPX Neo prototype itself is not crucial; what matters is whether it can provide StepFun with a real device bargaining chip when negotiating interfaces, permissions, and revenue-sharing rules with mobile phone manufacturers and App platforms.

Following closely behind is OpenAI. According to reports, its first AI Agent-enabled smartphone is expected to enter mass production as early as the first half of 2027. This arms race for AI agent terminals is never just a ranking contest among domestic players.

This scenario bears a strong resemblance to Google's 2007 strategy of forming the Open Handset Alliance with HTC, Motorola, Qualcomm, and operators to promote Android using real devices.

However, AI agent smartphones face greater resistance. Android helped mobile phone manufacturers reduce system development costs back then, benefiting nearly all participants. Once AI agents truly integrate into the system, they may alter the mobile phone manufacturers' own system entry points, bypassing the pages, ads, and recommendation chains that super Apps rely on for survival. No one can be certain they will be a pure beneficiary.

Therefore, for StepFun and ByteDance, the short-term goal may not be the number of units sold. What they need is a tangible real device to prove that their models can truly integrate into the operating system. Then, armed with this bargaining chip, they can renegotiate permissions, interfaces, and revenue-sharing rules with mobile phone manufacturers and App platforms. The smartphone is the most visible prop in this game, but the real negotiating table is not at the product launch event.

· Data Sources and Notes: QuestMobile's "Insight Report on AI Application Market Development in the First Half of 2026," official releases and public reports from StepFun, Honor, and ZTE/Nubia.

- END -