In ByteDance's own demo footage, nobody touches the phone. A voice asks it to compare one product across the shopping apps and order the cheapest, and the screen starts moving by itself: Taobao opens, a search field fills, JD.com launches, Pinduoduo loads.
Within two days, people who owned that phone were being logged out of WeChat.
That swing — a sell-out followed by a wall of refusals — is the real story of ByteDance's hardware move, and it has very little to do with silicon.
Key takeaways
- ByteDance did not build a phone. It placed its Doubao phone assistant (豆包手机助手) at the operating-system level of a ZTE nubia handset, the M153, launched on 1 December 2025 at 3,499 yuan (≈ US$490 / HK$3,815).
- The engineering prototype (工程样机) sold out on day one, then ran into resistance: WeChat logins failed, Taobao demanded human verification, and several banking apps treated the device as a risk environment.
- ByteDance retreated quickly, dropping WeChat operation on 3 December 2025 and then restricting itself inside financial and gaming apps.
- A consumer model, the nubia NaviX Ultra, went on sale on 16 September 2026 from 5,999 yuan (≈ US$845 / HK$6,540), or 5,499 yuan (≈ US$775 / HK$5,995) after the national subsidy (国补), carrying a new opt-in rulebook called SAEP.
- When SAEP's 30-day notice period ended on 14 October 2026, the published refusal list ran to roughly 155 apps — WeChat, QQ, Taobao, Alipay, Meituan, Kuaishou, Xiaohongshu, Pinduoduo and Didi among them.
What actually shipped
The division of labour was unusually clean. ByteDance defined the assistant and the experience; ZTE's nubia brand did the hardware design and manufacturing. The result was explicitly labelled a technical preview: an engineering prototype for developers and enthusiasts, sold with a warning that camera tuning and other software polish would lag mainstream flagships, and updated roughly every two weeks through the end of the first quarter of 2026.
ByteDance has been consistent on one point since day one: it says it has no plan to make its own phone, and that it is negotiating assistant deals with several handset makers. That is the strategic tell. A model company with no factories wanted the deepest possible distribution slot — the operating system — without paying for a supply chain, inventory or an after-sales network.
At 3,499 yuan it was cheap enough to be an impulse buy, and scarce enough that resale prices briefly ran well above list.
The pushback, and how fast it came
The demo showed the assistant working inside rival platforms — Taobao, JD.com, Pinduoduo, Dianping, Xiaohongshu, Trip.com — by reading the screen and simulating taps, rather than waiting for each app to expose an interface.
That is precisely the capability the app owners objected to. Users reported WeChat sessions terminated for "abnormal login environment"; Taobao raised verification challenges; banking apps blocked logins. Tencent has denied blocking WeChat. ByteDance did not fight: it removed WeChat control within days, and then restricted three scenarios — score farming, financial apps and competitive games — citing fairness and security.
The reasoning on the other side is not mysterious. If an assistant completes the task, the user never opens the app, and an app economy built on openings, impressions and in-app ad inventory loses the unit that pays for it.
The rulebook: from system privilege to declared consent
The first device leaned on a deep Android system permission that effectively makes an app part of the operating system. The second device took a different route. ByteDance published the Screen Automation Execution Protocol (SAEP), under which each app declares its own boundary: refuse screen automation entirely, or allow it while blocking specific actions such as posting, deleting, check-in rewards or claiming benefits. The protocol came with a 30-day public notice period that ended on 14 October 2026.
What emerged was less a negotiation than a vote, and the result was lopsided. Around 155 apps — the counts in reports range from 151 to 155 — declined, and they were not obscure utilities. They were the apps people open most. During the notice period, the assistant could mainly operate system functions and a small number of apps that had granted permission in advance: alarms, calendar entries, calls, local file search, recording summaries, screen reading.
Why this is a company story, not a gadget story
Two numbers explain the stakes. Chinese phone makers' own built-in assistants had already reached roughly 535 million users by September 2025, up 65 million in a year, against about 287 million users for standalone AI apps from internet companies — a gap measured by QuestMobile and cited in Caixin's reporting. The assistants were winning, and they were winning inside devices that the model companies do not control.
Ni Fei, president of ZTE's nubia Technology, framed the partnership as a bet on openness, comparing it to combinations like "iPhone + ChatGPT" and "Samsung + Gemini". His company needed the attention: IDC data cited by 36Kr put ZTE at 0.3% of China's smartphone market in the second quarter of 2026, ninth place, in a market where the top six brands hold roughly 96%.
According to Li An, chief scientist at BrainNet (脑机网), the struggle here is not about model capability but about who owns the interface layer, and interface ownership is settled by permissions and contracts rather than by benchmark scores.
Honest limitations
- Neither company has published unit sales. The sell-out claim rests on company statements; the NaviX Ultra's reported 370,000 JD.com reservations come from media, not disclosure.
- The refusal-list count varies across reports between 151 and 155 apps, so I write "roughly 155" rather than treat one figure as exact.
- The causal chain for login failures rests on user reports and company statements. Tencent has denied blocking WeChat and no regulator has ruled, so treat "apps blocked it" as an allegation pattern, not an adjudicated fact.
- I have not examined Doubao's wider app business, ByteDance's own chip plans, or the unsettled legal status of OS-level screen automation under Chinese rules — each deserves its own piece.
- Prices exclude channel discounts and vary by configuration; the US dollar and Hong Kong dollar conversions are approximate.
What readers can do now
- Before buying any "agent phone," read the published opt-out list, not the spec sheet. The assistant's usefulness is bounded exactly by which apps said yes, and in this case the apps people use most said no.
- If you build or run an app, decide your automation stance now and declare it. Under a SAEP-style regime, silence is a decision; writing down what an agent may and may not do is cheaper than discovering it after launch.
- If you track the sector, watch the next OEM agreement, not the next model release. The scarce asset is a handset maker willing to grant system-level permission, and that is a partnership question rather than a technology one.