Claude Code 的走红反映了 AI 编程从“自动补全”向“代理式”的快速跃迁。2021 至 2024 年间，多数工具只能补几行代码；到 2025 年初，代理式产品出现。Anthropic 的 Claude Code 在这一时期推出，随着 Claude Opus 4.5 上线被认为达到拐点。负责人 Boris Cherny 表示，他个人写代码由最初的 5% 交给工具，提升到 Opus 4/4.5 时代的 100%；Anthropic 内部也高度采用，约一半销售团队每周使用，技术员工几乎 100% 使用，Claude Code 团队约 95% 的代码由该工具生成。
商业数据同样显示爆发式增长。2025 年 11 月，Claude Code 的年化经常性收入（ARR）达到 10 亿美元；到 2025 年底又至少增加 1 亿美元，占 Anthropic 约 90 亿美元总 ARR 的约 12%，成为增长最快的板块之一。Anthropic 目标在 2028 年实现正现金流。竞争对手亦受益于浪潮：Cursor 也在 2025 年 11 月达到 10 亿美元 ARR；OpenAI、Google 与 xAI 正加速推出各自的代理式编程产品。
Anthropic 正将动能扩展到非编程场景，推出可管理文件并与软件交互的代理 Cowork。Cherny 描述其日常会同时运行 5 到 10 个代理，跨终端、移动端与网页；Cowork 已被用于项目管理与自动提醒。公司预期，代理将在一年内覆盖更多“繁琐工作”，如填表、搬运数据和发邮件，释放人类时间。这种抽象层级的上升被视为编程史的延续，而非终结。
Claude Code’s rise reflects a rapid shift in AI coding from “autocomplete” to agentic systems. From 2021 to 2024, most tools could only suggest a few lines; by early 2025, agentic products emerged. Anthropic’s Claude Code launched in this window and is widely seen to have hit an inflection point with Claude Opus 4.5. Head Boris Cherny says his own coding went from 5% assisted to 100% with Opus 4/4.5; internally, about half of the sales team uses it weekly, nearly 100% of technical staff use it, and roughly 95% of the Claude Code team’s code is generated by the tool.
The business metrics show similar momentum. In November 2025, Claude Code reached $1 billion in annualized recurring revenue (ARR); by year-end it added at least another $100 million, accounting for about 12% of Anthropic’s roughly $9 billion total ARR and ranking among its fastest-growing segments. Anthropic aims to be cash-flow positive by 2028. Rivals are also benefiting: Cursor hit $1 billion ARR in November 2025, while OpenAI, Google, and xAI are racing to launch their own agentic coding offerings.
Anthropic is extending this momentum beyond coding with Cowork, an agent that manages files and interacts with software. Cherny reports running 5 to 10 agents at once across terminal, mobile, and web, using Cowork for project management and automated reminders. The company expects agents to handle more “tedious work” within a year—forms, data transfers, emails—freeing human time. This rise in abstraction is framed as a continuation of programming’s evolution, not its end.