Anthropic adds experimental Code Mode to Claude Code
Anthropic added Code Mode as a default-disabled experimental feature in Claude Code v2.1.323 and later. It lets Claude orchestrate multiple tool calls, conditions, and loops in local JavaScript, returning only selected results to the model. Third-party tests cited by 36氪 found API costs fell 22%–23% on two development tasks and cumulative processed tokens dropped 62%–77% in tool-heavy workflows, though some file tasks used more tokens and Haiku 5.5 produced three failed checks.
Code Mode changes how Claude Code orchestrates tools without changing the underlying model, and independent tests reported substantial savings for suitable multi-step workloads while also exposing task-dependent limits.
用过 Claude Code 的开发者,几乎都有过同一种体验:干活确实爽,但看着后台飞速跳动的 Token 消耗和账单,心也在滴血。
过去想让 AI 智能体跑得更快、花钱更少,用户通常只有两条路:要么苦等厂商发布更便宜的新模型,要么手动把主力模型降级成小模型,忍受交付质量的滑坡。
但最近,Anthropic 在 Claude Code(v2.1.323 及以上版本)的更新日志里,悄悄塞进了一个默认关闭的实验功能——Code Mode。第三方实测显示:在完全不换模型、不牺牲智商的前提下,开启这个模式跑多步开发任务,真实 API 账单直接少了 22%–23%;而在高频调用外部工具的场景下,模型累计处理的 Token 总量更是暴跌了 62%–77%。
#Anthropic#Claude Code#Code Mode#Opus 5.5#Sonnet 5.5#Haiku 5.5#Fable 5.2#Bicameral#bestagent.dev