The landscape of agentic artificial intelligence is shifting rapidly, with two major announcements this week from DeepSeek and Microsoft Research that challenge the dominance of established players like OpenAI and Google. While the news is grounded in the tech sector, its implications for broad market sentiment and developer ecosystems cannot be ignored.
DeepSeek, the Chinese AI lab backed by High-Flyer, is forming a dedicated “Code Harness” team in Beijing to build a native agentic coding tool—a direct competitor to Anthropic’s Claude Code and OpenAI’s Codex. Engineer Deli Chen confirmed the hiring drive on X, seeking a product manager and an R&D engineer with hands-on experience across every major coding agent: Claude Code, Codex, Cursor, GitHub Copilot, Manus, and OpenClaw. The internal formula is “Model + Harness = Agent,” signaling a full-stack ambition beyond just supplying the underlying model. DeepSeek’s latest V4 Flash model is priced at $0.14 per million input tokens, up to 100x cheaper than Anthropic’s Claude Opus 4.7, giving any in-house harness a structural cost advantage that could reshape developer workflows.
Meanwhile, Microsoft Research has released Fara1.5, an open-weight family of browser-use AI models that outperforms both OpenAI Operator and Google Gemini 2.5 Computer Use on key benchmarks. On Online-Mind2Web, the 27-billion-parameter model scored 72%, compared to 58.3% for OpenAI and 57.3% for Google. Even the 9B version surpassed both with 63.4%. The models were trained using a novel pipeline that leveraged GPT-5.4 as a teacher agent, and they run locally with a strong safety layer via MagenticLite, a sandboxed browser environment that logs every action and allows user intervention. The 9B variant is live now on Azure AI Foundry, with 4B and 27B following shortly, and Microsoft plans to expand beyond the browser into desktop and enterprise software.
