Wink Pings

Anthropic Just Gave Claude Code a New Mode: Three Models Split the Work, Set Up in 1 Minute

Claude Code's Code Mode divides work among models: Opus 5.5 plans, Sonnet 5.5 executes, and Haiku 5.5 grades. Setup takes one minute and one environment variable. This article covers the rationale, the steps, and one easy-to-miss pitfall.

Anthropic has just given Claude Code a new trick, officially called Code Mode. Setup takes one minute and one environment variable.

Claude Code is Anthropic's command-line coding assistant. Normally, to get work done, it has to call tools such as Bash, Read, and Edit one after another. Code Mode changes this process: Claude writes a piece of JavaScript, runs all tool calls in parallel, and returns only the part of the results that needs to be read. This eliminates the back-and-forth of opening 30 files. While a long-running task is in progress, Claude can keep responding to you.

Models are divided into three tiers:

  • Opus 5.5 only does planning and writes code blocks.
  • Sonnet 5.5 executes builds in subagents.
  • Haiku 5.5 grades against a rubric and appears at three checkpoints: after each build, after 3 consecutive rounds of failed checks, and before finishing.

Anthropic's own numbers: an independent grader can increase task success rate by up to 10 percentage points; if checks are still red for 3 consecutive rounds, it stops and comes to you instead of continuing to burn tokens.

shiraha posted a demo on X:

Setup method:

  1. Upgrade to Claude Code v2.1.323 or later. agent teams requires v2.1.326.
  2. Paste the prompt from the official Code Mode documentation. It will read the official docs first, then act.
  3. In the env in user settings, add "CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1", set the main model to Opus 5.5, use Sonnet 5.5 as the builder for subagents, and Haiku 5.5 as the grader.
  4. It first backs up the configuration, then shows all changes via diff; they take effect only after you confirm.
  5. If there is no effect after configuration, run /status first.

Pitfall: Code Mode only runs in an interactive terminal. -p, the VS Code panel, and Bedrock and Google Cloud sessions do not use this mode.

Behind this is a more general approach. rari (@0xwhrrari) summarized it as mode engineering:

  • Judgment-type: humans need to look at it, mistakes are costly, use the strongest model.
  • Routine-type: humans also look at it, but it looks the same every day, use a cheaper model.
  • Things no one reads as prose: tagging, ranking, filling fields, use the cheapest model.
  • Things with a single correct answer: summing, checking dates, determining "whether there is a change," write a script and don't run a model.
  • Nothing to do: don't run it.

Claude Code's Code Mode is an official implementation of this approach. For the remaining two parts, use scripts where scripts are needed. For monitoring that runs once an hour and finds nothing 23 times a day, the most important thing to add is a precondition: if there is no change, don't run it.

发布时间: 2026-11-21 02:36