WAIC 2026 Smartphone Roundup: Three AI Phones Take Three Routes — One Rebuilds the OS, One Connects Apps, One Mounts a Robotic Arm
Recapping the smartphones at WAIC 2026, three AI phones take three different routes: one rebuilds the system, one connects apps, and one goes straight for a robotic arm. On July 20, the 2026 World Artificial Intelligence Conference wrapped up in Shanghai. Over 1,100 companies exhibited, with 349 products making their global debut. But if there's one track that was the most fiercely competitive this year, agentic phones definitely take the top spot. Three products shared the stage, each touting a different slogan—and as a reporter, I'm more than just an observer.
💡 What You Will Learn
Recapping the smartphones at WAIC 2026, three AI phones take three different routes: one rebuilds the system, one connects apps, and one goes straight for a robotic arm. On July 20, the 2026 World Art
📜 Table of Contents
WAIC 2026 Smartphones Roundup: Three AI Phones, Three Different Paths — One Rebuilds the OS, One Connects Apps, One Adds a Robotic Arm
On July 20, the 2026 World Artificial Intelligence Conference wrapped up in Shanghai.
Over 1,100 companies exhibited, with 349 products making their global debut. But if there's one track that was most fiercely competitive this year, it's agentic phones. Three products shared the stage, each with its own slogan — and there were more journalists than attendees, all crowding around the booths to film the phones.
But more interesting than the buzzwords are the three completely different paths they've taken.
StepFun: Rebuilding the Operating System from Scratch, Making AI a "Native Resident"
StepFun's STEPX Neo took home WAIC's "Crown Jewel" award.
Its approach is the most radical — instead of adding an AI assistant to Android, it rebuilt an entire operating system from the ground up, called Step AOS.
Yin Qi put it this way: if you open a door for an agent on an old system, it will always be a guest; but if you build a house for the agent, it becomes a native resident.
Step AOS does three things at the system level: it centrally allocates chip compute power, reorganizes user data into a format AI can understand, and breaks down phone capabilities into the smallest components, letting agents freely combine and invoke them like building blocks.
A live demo showed this scenario: the user says, "Help me prepare for tomorrow's presentation," and the built-in agent Amoo automatically reads Feishu calendar entries, checks unread messages, organizes content outlines, generates a PPT, and sends it — all without any manual operation. Amoo also remembers your preferences — if you favor tech-blue color schemes, it will generate visuals in tech blue.
The STEPX Neo is currently the only agentic phone to pass the national L3 intelligence grading test (the highest level available). But you'll have to wait until late October to buy one — it's still in the production pipeline.
Nubia: No OS Rebuild — Using MCP Protocol to Connect Apps
Nubia's NaviX Ultra took a different route — instead of rebuilding the OS, it uses MCP and A2A protocols to enable direct, underlying interoperability between the phone's AI and each app's agents.
Last December, Nubia partnered with ByteDance to launch the first-generation Doubao Phone M153, which used a "simulated click" approach — the AI would open apps one by one in the background, find buttons, fill in forms, and operate the phone like a human. But whenever an app updated its UI or moved a pop-up, the task chain could break — and privacy restrictions from major apps also limited its permissions.
The second-generation NaviX Ultra changed course: instead of simulating clicks, it lets apps provide their own MCP service interfaces, opening up select data and control permissions. You say, "Book a hotel for me," and the AI directly calls Ctrip's agent interface to complete the search and reservation — no need to "pretend to be human" by tapping the screen.
Nubia President Ni Fei summarized the core capabilities of agents as "understanding what you say, getting things done, remembering your preferences, and staying secure." The NaviX Ultra is positioned as a mass-production flagship, and its on-device large model completed filing on July 15 — but only a prototype was shown on-site, with no hands-on experience available.
Honor: Giving the Phone a Robotic Arm, Letting AI Move in the Physical World
Honor's Robot Phone follows a path completely different from the other two — while they let AI operate apps inside the phone, Honor wants AI to operate the physical world too.
The Robot Phone integrates a four-degree-of-freedom titanium alloy gimbal into the top of the device, stored inside the body during normal use and popping out in 0.8 seconds when needed. It supports 360-degree rotation tracking and intelligent camera movement. Paired with a 200MP gimbal main camera and Log-C encoding developed in partnership with ARRI, the phone can automatically track subjects, adjust composition, and execute camera moves — letting one person shoot cinematic footage solo.
Under the hood, it's powered by the Magic agent large-model matrix, co-developed by Honor and Alibaba, with voice, gestures, eye contact, and body movements all serving as inputs. A live demo showed a cross-app task: "Find the hotpot restaurant I saved on Douyin and use Didi to take me there directly" — the AI automatically completed everything from search to navigation, only asking the user to confirm at the final payment step.
The Robot Phone is now open for pre-orders across all channels, powered by the Snapdragon 8 Gen 5 Ultimate Edition, and is slated for a Q3 launch.
Three Phones, One Shared Future
Three phones, three paths — behind them lies the industry's differing views on what an AI phone should look like.
Wang Xiaoya, an analyst at Sigmaintell Consulting, believes the fundamental difference between AI phones and traditional phones lies in the shift of interaction models — moving from "app-centric" to "intent-centric."
But the challenges are just as obvious. A cross-app task with five steps, each with a 90% success rate, yields only a 59% overall success rate. Once more agents are deployed, performance drops significantly. Privacy and security, ecosystem collaboration, and commercial profit distribution — every one of these is a tough nut to crack.
Still, the direction is clear: in 2026, four conditions — on-device compute power, model capabilities, regulatory filing, and market demand — have aligned for the first time. The smartphone industry has been declining for five consecutive quarters, and everyone is desperate for a reason to make users upgrade.
AI agentic phones might just be that reason.
These three paths — system-level reconstruction, app-level integration, and physical-world extension — which one do you think will win? Let's discuss in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
