HomeArticle

Regarding the second-generation Doubao mobile phone, these are the 10 questions I care most about.

唐辰同学2026-09-20 08:01
Doubao has learned to "knock on the door" and no longer barges in rudely.

On September 16, the Nubia NaviX Ultra (hereinafter referred to as the Doubao Phone 2nd Gen) officially went on sale, with a starting price of 5,999 yuan. The new device is equipped with the Doubao Mobile Assistant (Consumer Edition). It is defined as the first AI agent phone, marking that this category has moved from engineering prototypes to mass production and commercial deployment.

In December 2025, the technical preview version Nubia M153, namely the Doubao Phone 1st Gen, was released in limited quantities at a price of 3,499 yuan. The first batch of 30,000 units was sold out quickly, once driving ZTE's share price to hit the daily limit.

Its core value lies in that it provides a feasible path for AI phones to achieve true "deep intelligence", and it is also the first large-scale deployment of GUI Agent (Graphical User Interface Agent) on real consumer devices.

But just a few days later, the Doubao Phone 1st Gen hit the "ecosystem wall" of super apps. For example, abnormal login issues occurred on WeChat, and risk control warnings were continuously issued by apps such as Taobao and banking applications. Doubao had to disable financial and some sensitive operations, and the "fully automatic" phone was eventually called off.

Regarding the Doubao Phone 2nd Gen, the 10 questions below are what I care most about. After reviewing various discussions and information, I found the answers I was looking for.

The content is a bit long, you can follow the subheadings to read the topics you are interested in.

01 How many flaws of the 1st Gen have been fixed in the 2nd Gen?

The core controversy of the 1st Gen was that the Doubao Mobile Assistant looked too much like a third-party plug-in. Starting from December 3, 2025, when users used Doubao to operate WeChat on the M153, they were prompted with "abnormal login environment" and even kicked offline. WeChat responded that "there is no special action, it may have triggered the existing built-in security risk control measures".

Doubao reacted very quickly, and took down the "WeChat operation" capability that same night. Subsequent technical reviews pointed to the system-level permission INJECT_EVENTS, combined with screen reading and OCR, which was identified by WeChat as automated behavior of scripts/plug-ins.

The 2nd Gen clarifies the legitimacy of its "identity". First, the Nubia Doubao mobile large model has completed the filing of end-side generative AI services with peer vendors including Apple Intelligence and Huawei Celia;

Second, Doubao launched the SAEP (Screen Automated Operation Declaration Protocol) before the new phone went on sale: it will be public for 30 days, no operation will be performed if the user does not agree within the period, and access is granted by default if no rejection is received after the period expires; it is also equipped with operation log query, so that third parties can trace the operation decisions of the assistant. From operating covertly to asking for permission in advance, this is a difference in nature.

The 2nd Gen also makes up for the deficiencies of the 1st Gen engineering prototype in terms of completion. It is a complete retail product with network access certification, 24-month interest-free installment, and 180-day free replacement without repair, and its hardware is also upgraded to flagship level. As for whether the starting price of 5,999 yuan is worth it, every user will have a fair evaluation in mind.

In addition, the biggest difference between the 2nd Gen and the 1st Gen lies in the execution strategy, which has changed from direct screen reading plus simulated click to protocol priority, simulation as fallback, and manual confirmation. The main line turns to MCP (Model Context Protocol) and A2A (Agent to Agent), GUI screen reading is reduced to a backup measure, and third-party apps can declare boundaries from three levels: page, business intention, and operation type.

This path reaches the same destination as mobile phone manufacturers such as Apple, Huawei, Honor and Xiaomi, all trying to bypass direct operation of app interfaces and turn to system interfaces and Agent collaboration.

But this also creates a new problem: the problem of the 1st Gen is "can I do it", while the problem of the 2nd Gen is "will others allow me to do it". This will bring a confusing perception to consumers: will the Doubao phone with a changed technical route still be the Doubao phone?

There are also several remaining issues that the 2nd Gen has not given answers to.

There are also some flaws that the 2nd Gen has not fixed. The 1st Gen clearly stated in the "Notice on Adjusting the AI Mobile Operation Capability": ① Restrict scenarios of point brushing/incentive brushing; ② Temporarily take down the operation capability of financial apps such as banks and internet payment platforms; ③ Restrict AI capabilities in game scenarios involving competitive rankings.

Perhaps these restrictions will not be lifted in the short term. Some media tested the 2nd Gen on its launch day: the voice operation Agent can access ByteDance's full range of products and a small number of third-party apps such as Cao Cao Travel, while external e-commerce apps such as Taobao Flash Purchase and JD do not support related operations for the time being;

Other media reports said that WeChat only opens part of its capabilities (sending messages via voice, audio and video calls), payment and transfer still require dual authentication, and Taobao, Meituan, Alipay and banking apps do not support direct AI operation for the time being.

The "Mobile Operation" support range listed on Nubia's official website also confirms this: system and ZTE applications, ByteDance series applications, and some accessed third-party applications (Cao Cao Travel, maps, Qishui Music, Feishu Minutes).

The 2nd Gen also "buries a hidden risk" in Token calls: AI operations are pointed out to have a limited number of free times per day, and additional payment is required for unlimited use. Some reviews pointed out that the 2nd Gen "has a limit on the number of free uses of AI operation functions per day, and additional payment is required for unlimited use". If this review is confirmed by the official, it means that the 2nd Gen will have a fatal experience breakpoint, as it turns its core selling point into a billing item.

02 After mass production, what does it mean for Doubao, ZTE and the AI mobile track respectively?

For Doubao, this is four accounts to calculate:

The first layer: system-level identity. In the past, Doubao was just one of dozens of apps in the mobile phone, with permissions limited to "answering questions". On the 2nd Gen, Doubao is a system role resident on the desktop, with an exclusive physical AI key integrated with fingerprint authentication. Even the cloud service account system is carried by the Doubao Mobile Assistant account instead of the Nubia account. This is an identity upgrade beyond functions.

The second layer: real task data. This is what I think is the most underrated point. Conversations inside the app can only see behaviors within the wall; system-level Agent can see the complete cross-app decision chain, such as what the user wants to buy, which platforms they compare prices on, and why they give up. As the AI application with the largest user volume in China, what Doubao lacks is exactly the real "user behavior link" data, which is the fuel for the next generation of algorithm recommendation and advertising.

The third layer: turning Token from "pricing information" to "pricing transaction". Once the Agent can be implemented as an order, Doubao's monetization unit will no longer be subscription fee or CPM advertising fee, but commission. For example, there was once widespread industry rumor that Doubao has piloted commission deduction for life services, with a comprehensive hotel rate of 12%. Although the official clarified it, it did not respond to the substantive issue.

The fourth layer: execution capability expands from "inside the screen" to "the physical world". This is the most easily overlooked ceiling of Doubao. No matter how smart the GUI Agent is, what it can move is always pixels: it can click buttons, but cannot move any physical object.

Doubao has begun to expand from mobile phone screens to automobiles. On September 17, the Doubao Cockpit Assistant was released, which can perceive hundreds of vehicle signals, and link thousands of vehicle-level atomic capabilities such as vehicle control, navigation, assisted driving, and entertainment. When the user says "take a nap, wake me up at 1 o'clock", it will automatically adjust the seat, close the sunshade, reduce the air volume, play soft music, turn on massage, wake the user up on time and remind the schedule.

If you connect these four layers, what Doubao wants is not just the "AI mobile entrance", but a cross-device execution and memory layer that follows the user around.

For ZTE, the value of brand breakthrough is far greater than the financial value on the books.

The consumer business of ZTE, to which Nubia Mobile belongs, achieved a revenue of 33.816 billion yuan in 2025, a year-on-year increase of 4.35%. In the first half of 2026, the consumer business reached 18.629 billion yuan, a year-on-year increase of 8.09%.

In the caliber of third-party institutions such as IDC, Nubia does not disclose its sales volume separately, and is included in the "others" category. Therefore, the significance of the Doubao Phone 2nd Gen to ZTE is that it complements the consumer-side AI terminal puzzle for the narrative of "Connection + Computing Power + AI for All"; at the same time, it can obtain an industry-level breakthrough opportunity with extremely low independent R&D risks.

But ZTE must be sober that even if it is calculated based on the rumored 200,000 units of stock preparation and an average price of about 6,000 yuan, the total revenue is only about 12 billion yuan. The attitude of the capital market, from the daily limit of the 1st Gen to the slight rise of the share price of the 2nd Gen, shows that the emotional premium brought by the Doubao Phone has been overdrawn by the 1st Gen.

Ni Fei, President of Nubia, is very calm. He said, "The first half of the AI mobile era is adding AI functions to mobile phones, and the second half is enabling mobile phones to do things for people through AI agents" and "2026 will definitely be the first year of the popularization of AI agent phones".

But for the AI mobile industry, the value brought by the Doubao Phone may not be positive. The 1st Gen tried to prove that GUI Agent can be done, but it used a "collective block" to put the attitude of super apps towards "barbaric intruders" on the table, and raised the wall. In the words of a commentator, what the Doubao 1st Gen hit was "the ecological wall built by internet giants for decades".

The 2nd Gen learned the lesson and took a step back, hoping to get through the process of "applying to enter the super app fence". That is to say, SAEP is essentially an attempt to transform "technical conflict" into "rule negotiation". Many commentators compare it to the Robots protocol in the AI era, whose value is higher than the price tag of the mobile phone itself.

03 Is now the best window period for the Doubao Phone? How long will this window last?

Theoretically yes.

First, the regulatory window is open. On July 15, the Cyberspace Administration of China announced that 7 end-side AI services have completed the filing; the national standard for intelligent classification has been implemented, and about 63 models in the first batch have reached L3; industry associations have released the agent protocol requirements compatible with MCP and A2A.

Second, the competitor's path has not been realized yet, and all are stuck at the same threshold. Honor is currently the most active brand promoting AI phones. Sun Jianfa, President of MagicOS, clearly stated that although the automatic execution capability has been developed and verified, it is still promoting regulatory certification;

Apple's services in China are in an incomplete state. Although Apple Intelligence has been filed, the reconstructed Siri has not been launched in China so far. It is very rare that everyone is standing on the same starting line waiting for the starting gun.

Third, the cognitive window has been opened for free by the 1st Gen. 3499 yuan, 30,000 units, second-hand premium, and the business of renting Doubao phones for 50 to 79 yuan a day, these do not require the 2nd Gen to spend a penny on marketing to explain what "AI helps you operate the mobile phone" means.

But there are also four unfavorable conditions that are compressing the window period of the Doubao Phone 2nd Gen.

First, installation does not equal usage. QuestMobile data shows that in June this year, the MAU of mobile phone manufacturers' AI applications was 755 million, a year-on-year increase of only 14.0%, with 51.4 times of monthly usage per capita; in the same period, the MAU of AI native apps was 499 million, a year-on-year increase of 85.4%, with 92.7 times of monthly usage per capita.

That is to say, the activity of the "pre-installed traffic pool" of AI phones is much lower than that of "actively downloaded native apps". Another survey shows that more than 70% of users turn off most AI functions half a year after purchasing an AI phone.

Second, the storage super cycle is squeezing the popularization of AI phones. TrendForce data shows that since 2026, the contract prices of conventional mobile DRAM and NAND have risen sharply. For example, in the first quarter of this year, they rose by 90%-95% and 55%-60% month on month respectively. The proportion of storage in mobile BOM has risen from 10%-15% to 30%-40%, making it the most expensive component of the whole device. As a result, AI capabilities are forced to be stacked on high-end models, while popularization requires mid-range products.

Third, super apps have opened the door for everyone, except for Doubao. Taking WeChat as an example, in July this year, Tencent customer service confirmed to *Caijing* that WeChat is cooperating with Huawei, Honor, Xiaomi, OPPO, and VIVO to launch A2A assistant capabilities, and Honor has completed the adaptation in the first batch. But Doubao is not on this list of approved partners.

Many Tencent insiders said that Tencent has always advocated A2A instead of GUI, on the grounds that it is more secure and compliant. This is obviously taking sides. ByteDance's traffic is more of a threat than a cooperation to Tencent.

Fourth, the right to define is being snatched by others. Some analysts believe that "2026 can be regarded as the first year of agent phones"; IDC predicts that China's new generation of AI mobile shipments will reach 147 million units in 2026, accounting for 53% of the overall market. But these are all predictions, not the real performance of AI phones. Moreover, the realization of this goal depends on the entire mobile phone camp, not just Doubao.

Based on this, my judgment is that what the Doubao Phone is fighting for is a 12-18 month "definition right window", not a "sales volume window". If the mobile phone industry releases the following three signals, it means the window for the Doubao Phone is closed:

Leading manufacturers complete the closed loop of "self-developed end-side model + self-owned Harness" and no longer need external agents;

Super apps and leading manufacturers reach a mature plan for computing power bills and commission sharing, and lock in the cooperation list;

The storage cycle falls, so that configurations above 12GB are reintroduced into mid-range price points.

According to the current pace: Honor has fully pushed WeChat A2A, Huawei HMAF 2.0 is open, and Xiaomi has deployed multi-model scheduling at the same time; I tend to think that the window period of the Doubao Phone will narrow significantly in the second half of 2027. The reason is not that Doubao's technology is not good, but that the preparations that various mobile phone manufacturers need to complete are almost finished.

02 After mass production, what does it mean for Doubao, ZTE and the AI mobile track respectively?

If it is only to compete for the AI mobile entrance, the Doubao Phone 1st Gen has already completed the test.

If you put cars and mobile phones together, you will find a detail that is easy to ignore: SAEP is not needed in cars.

The in-vehicle space is equivalent to the owner standing at the door to welcome guests, and more than 2000 SOA interfaces are actively called out by the vehicle system. But on mobile phones, Doubao has to draft a protocol, explain to each party that it will not tamper with your content, and configure 30-day public notice, rejection taking effect immediately, and traceable logs. The more comprehensive the protocol is written, the more it shows that what it has in this mobile phone is borrowed.

What it really wants is to turn the borrowed execution environment into a default protocol layer with its own execution environment. ByteDance needs to obtain the system rules of Agent mobile phones through SAEP, clarifying whose behavior needs to be declared, to whom, and how long the declaration takes effect.