Chinese Startup StepFun Launches a Phone Built Around an AI Agent, Not Apps

What it claims

StepFun's STEPX Neo (3,999 yuan, about $560) ships Step AOS 1.0, which StepFun says was built from scratch around its agent, Amoo, rather than adding an assistant to iOS or Android the way Siri and Gemini are. Long-term memory, task scheduling and security controls sit in the base layer. StepFun describes the OS as a mix of Android, Linux and real-time OS components. The price bundles 12 months of the STEPX Token Plan at 100 billion tokens a month. StepFun has not said what AI access costs after the first year.

How it works

Phone functions (messaging, apps, files, settings) are exposed as MCP-callable blocks. Users get two entry points: a conventional app grid and a separate agent space. A second MiniLED screen on the back shows what Amoo is doing.

Requests are routed by difficulty, and transactions stay with partner platforms:

flowchart TD
  U[User request] --> R{Simple or multi-app?}
  R -->|Simple, e.g. alarm| D[Small on-device model<br/>some offline]
  R -->|Spans several apps| C[Larger cloud model]
  D --> M[MCP blocks: OS functions]
  C --> M
  C --> P[Partner platforms<br/>complete payment/booking]
  M --> A[Amoo memories:<br/>user facts + task learning]

The partner route is StepFun's answer to ByteDance's Doubao phone agent. Doubao drove apps by reading the screen and simulating clicks, and major Chinese apps blocked it in December. Nearly 100 partners have signed on to Amoo, including Alipay, Tencent, Ctrip, Didi and Meituan. Payments, travel bookings and content creation are still completed inside their own platforms.

How to use it

• Phone: pre-orders opened Oct. 13 on JD.com, Tmall and Douyin, with a first sales event on Oct. 16. The device was first shown in Shanghai on July 13, without a price.
• Models: Step 3.5 Flash, released earlier this year, is open-source under Apache 2.0, so it is the usable artifact for anyone building on StepFun's models. Step 5 Preview (600B parameters, 27B active, 1M-token context) arrived in September. The page gives no licence or access terms for it.

Limits

• The article gives no benchmarks, latency figures or task success rates for Amoo.
• It does not say which commands work offline or where the on-device/cloud split falls.
• It does not say whether third-party developers can expose their own MCP blocks.
• Post-year AI pricing is unannounced. Anyone counting on the device should treat the ongoing cost as unknown.
• Agent coverage depends on the partner list. Apps outside it get no stated path; the screen-reading fallback is not described.
• This is a news report built on launch-event claims and has not been independently tested.