Tencent-Backed StepFun Launches $560 STEPX Neo Phone Built Around an AI Agent

StepFun, a Tencent-backed Chinese artificial intelligence startup, has launched the STEPX Neo, a smartphone with an AI agent embedded in the operating system rather than offered as a standalone app. Pre-orders opened at launch at 3,999 yuan, roughly $560, and the first sales event is scheduled for 16 October.

An agent in place of the app grid

The pitch is a change in how people get things done on a phone. Today, booking a ride means opening a ride-hailing app, entering a destination and choosing a vehicle. On the STEPX Neo, according to Decrypt, the user states where they want to go and the agent completes the remaining steps itself.

An agent, in this sense, is software given a goal and left to work out and execute the steps needed to reach it. That distinguishes it from a conventional voice assistant, which typically answers questions or triggers single actions.

StepFun framed the launch around that distinction. It told the audience that in what it calls the agent era, AI stops being an add-on to the smartphone and becomes the product.

StepFun had previously presented its agent, operating system and STEPX branding at a Shanghai event held on 13 July, but gave no price or release date at the time. It argued then that conventional smartphone software had never been designed for autonomous task execution.

How Step AOS is built

The operating system, the core software layer that runs a device as iOS and Android do, is called Step AOS 1.0. StepFun contrasts it with Apple's Siri and Google's Gemini, which operate as assistants sitting inside their respective platforms.

Step AOS, the company says, was designed from the ground up around the agent, with long-term memory, task scheduling and security controls placed in the foundational layer. It is not entirely new code: StepFun describes it as a combination of Android, Linux and real-time operating system components.

The key architectural choice is to split phone functions such as messaging, apps, files and settings into small modules the agent can invoke directly. These modules are made available through MCP, short for Model Context Protocol, a common standard allowing AI systems to link up with other software. Users can still reach their apps through a familiar grid, but there is a separate space dedicated to the agent.

The agent itself is named Amoo. It keeps two memories: one holding the user's facts, habits and preferences, and another recording what it has learned about carrying out tasks.

Routine commands, such as setting an alarm, are handled by a small model on the device, in some cases offline. More complex requests that span several apps are sent to a larger model in the cloud. A second MiniLED display on the back of the handset shows what Amoo is doing.

Partners rather than simulated taps

Perhaps the most commercially significant detail is the partner list. Nearly 100 companies and products have signed up to work with Amoo, including Alipay, Tencent, Ctrip, Didi and Meituan. Payments, travel bookings and content creation are still completed on those firms' own platforms.

That model is StepFun's answer to the approach used by Doubao, which relied on reading the screen and simulating clicks to operate apps on a user's behalf. Screen-reading lets an agent work with apps whose owners have not cooperated, but it can be fragile when interfaces change, and it raises questions for app owners about control over their own transactions.

Formal integrations are, in principle, more dependable and give partners a say in how their services are used. The trade-off is reliance on those partners' willingness to take part.

Signing up some of China's best-known payments, travel, ride-hailing and delivery services at launch suggests StepFun has persuaded at least part of the domestic app ecosystem that agent-driven traffic is worth accommodating rather than resisting.

The model behind the phone

StepFun was founded in 2023 by Jiang Daxin, a former Microsoft vice president, and develops its own large language models, the class of AI system that underpins ChatGPT.

Its flagship, Step 5 Preview, was released in September. It has 600 billion parameters, the internal values a model adjusts during training, of which 27 billion are active at any one time. Activating only a fraction of a model's parameters per request is a common way to keep a very large system cheaper to run.

Step 5 Preview can accept up to 1 million tokens at once. Tokens are the chunks of text that AI services count and bill by, so a large context window matters for tasks that require holding a lot of information together, such as a user's activity across several apps.

A smaller model, Step 3.5 Flash, came out earlier this year and is freely available as open-source software, licensed under Apache 2.0.

Owning the model, the operating system and the hardware gives StepFun a degree of vertical control that few AI startups attempt. It also means the company carries the cost and risk of every layer.

Costs and open questions

Bundled into the 3,999-yuan price is a year of the STEPX Token Plan, granting a monthly allowance of 100 billion tokens. StepFun has not said what access to the AI will cost once that first year ends.

Because the more demanding tasks are processed in the cloud, serving them carries an ongoing expense. Buyers have no clear indication yet of whether the phone's central feature will come with a recurring fee after the bundled period.

There are other uncertainties. A phone whose main selling point is an agent acting on a user's behalf depends heavily on that agent being reliable; mistakes in bookings or payments carry more consequence than a wrong search result.

The long-term memory of personal habits and preferences also concentrates sensitive data, which puts weight on the security controls StepFun says are built into the system's base layer.

The partnership approach, while cleaner than screen-reading, limits Amoo to services that have signed up. Nearly 100 is a meaningful start, but well short of the full range of apps most users rely on.

Points to watch as the device reaches customers:

• How the phone performs once sales begin on 16 October, and early reports on Amoo's reliability
• Pricing for the token plan after the first 12 months
• Whether more app developers join the partner programme
• Whether established handset makers respond with deeper agent integration of their own

For now, the STEPX Neo is a test of a broader proposition: that the app-centred smartphone can be reorganised around software that acts on a request rather than waiting to be tapped.