DeepSeek Is Building Its Own Claude Code. Beijing Wants the Whole Stack

In brief

• DeepSeek is recruiting a "Code Harness" team in Beijing to build a native agentic coding tool—a direct rival to Anthropic's Claude Code and OpenAI's Codex.
• DeepSeek V4 already integrates natively with Claude Code, making this a move to own the full agentic stack rather than just supply the model underneath it.
• V4 Flash runs at $0.14 per million input tokens—up to 100x cheaper than Claude Opus 4.7—giving any DeepSeek-native harness an immediate, structural price advantage.

It started as a side project. Now it's coming for Anthropic's most prized product. DeepSeek, the Chinese AI lab born out of quantitative investment firm High-Flyer, is forming a new team to build "Code Harness"—an agentic coding tool that would compete directly with Anthropic's Claude Code and OpenAI's Codex. The announcement came from Deli Chen, a DeepSeek engineer who posted two open job listings on X this week: a product manager and an R&D engineer. Both jobs are based in Beijing.

Sadly for Western devs praising the remote work gospel, the company requires being based in the capital of China specifically. It’s not good enough to live in even Hangzhou, where High-Flyer, the hedge fund behind DeepSeek, was founded or Shenzhen, which is the Silicon Valley of China.  They want someone in Beijing—the symbolic and political capital of Chinese tech ambition, and the city where the government's relationship with its AI industry is the most direct, the most deliberate, and the most closely watched by Washington. "You can call it DeepSeek Code or something," Chen wrote in the post, adding laughing emojis as if the implications weren't serious at all.

They are.

> 🚀 We’re hiring! DeepSeek is forming a new Harness team to build Code Harness from the ground up—may be you can call it DeepSeek Code or something like this hhh🤣🤣🤣 > > 📍 Based in Beijing. Two roles open:
> 🧠 Harness Product Manager → https://t.co/vb3aWbYV9L
> 👨‍💻 Harness R&D… > > — Deli Chen (@victor207755822) May 20, 2026

The job listings, posted the same day on High-Flyer's recruitment platform, are unusually candid about what DeepSeek is building. The team's internal formula: "Model + Harness = Agent." The product manager role specifically requires candidates to have hands-on experience with Claude Code, Codex, Cursor, GitHub Copilot, Manus, and OpenClaw—essentially a who's-who of every agentic tool DeepSeek intends to compete with. Salary for both positions is undisclosed. Why tools matter as much as the models During the gold rush some became rich digging for gold, others got rich selling shovels. Deepseek wants to play both sides. Agentic coding tools—programs that don't just suggest a line of code but autonomously plan, write, test, and debug entire software projects—have become the fiercest battleground in AI. Anthropic's Claude Code is a command-line tool that lets developers hand off complex engineering tasks to an AI that keeps working without constant hand-holding. OpenAI launched Codex and most tech companies are now fighting to build the AI that sits inside every developer's terminal. DeepSeek wants in. And it has a peculiar advantage: Its latest model, DeepSeek V4, already runs natively inside Claude Code. The V4 lineup, released on April 24, comes in two flavors. V4 Flash—the workhorse—is built for speed and handles agentic tasks at $0.14 per million input tokens. (Tokens are the most basic unit of information handled by large language models.) V4 Pro, the more capable tier, runs at $0.435 per million tokens during an introductory promotion running through May 31. Claude Opus 4.7, Anthropic's flagship, runs at $15 per million input tokens. That’s a major budget difference for teams running continuous, loop-heavy agent pipelines. The Code Harness move is about owning the full stack. Building a harness means DeepSeek would control the interface developers actually see and interact with—the checkpoints, the terminal commands, the rollback features, the integrations. That's where user loyalty lives and likely where the money follows.

DeepSeek has made a habit of underestimating itself loudly, then delivering. When it released R1 in January 2025, the reasoning model wiped nearly $600 billion from Nvidia's market cap in a single day—because it matched OpenAI's o1 at a fraction of the cost. Analysts called it a fluke. DeepSeek kept shipping. No launch date has been announced—but the job information is there in case you’re interested, speak Chinese, and are willing to work out of Beijing.