What actually shipped

Code Mode arrived the way consequential developer features often do now: one line in a changelog and a new documentation page, with no launch post and no benchmark. Version 2.1.323, dated November 12, lists it as an experimental opt-in, as reported by i10X News and AgenticFeed from the changelog. The documentation itself labels the feature experimental.

The mechanism is a genuine change in how the agent works, not a settings tweak. In a normal session, per the docs, Claude requests a tool call such as Bash, Read or Edit and reads each result before deciding the next call. In Code Mode, Claude expresses the orchestration in JavaScript instead. A single script can fire independent commands in parallel, pass one call's output into the next, and return only the slice of output it needs to reason about. The calls and diffs still appear in the conversation; the code itself sits behind a collapsed row in the transcript viewer, opened with Ctrl+O.

Where the 37% came from

The number now traveling with the feature is specific, which is why it spreads: 37% fewer tokens. It is also not a Code Mode measurement. Anthropic's Code Mode page gives no saving of any kind, in tokens, time or money, and the sources reviewed provide no Code Mode measurement validating the comparison.

The figure comes from a post published a year earlier, in November 2025, introducing advanced tool use on the developer platform. That post describes Programmatic Tool Calling, an API feature in which Claude writes Python that runs in a sandbox, and reports that on complex research tasks average usage fell from 43,588 tokens to 27,297, a 37% reduction. Similar idea, different product surface, different runtime, different task set. Porting it across assumes the saving survives the move from an API sandbox to an interactive terminal session, and nothing in the sources reviewed tests that assumption.

This does not make Code Mode useless. Batching calls and filtering output before it reaches the model should reduce overhead on call-heavy work; that is the design, and it is plausible. What readers do not have yet is a number for this feature, from Anthropic or from an independent test. The honest version of the story is that the saving is unmeasured for Code Mode, not that it is 37%.

The conditions are the story for teams

The requirement list is long, and it decides who can even try this. Terminal only, native build only, a Claude subscription or Console account: no Bedrock, no Claude Platform on AWS, no Google Cloud's Agent Platform, no Foundry, no gateways. No VS Code panel, no Desktop app, no cloud sessions, no Agent SDK. Organizations on the HIPAA configuration or zero data retention are excluded. A session started in Code Mode can only be resumed in Code Mode.

The permission design deserves a slower read than it has had. Per-call deny and ask rules and PreToolUse hooks still see every call the script makes. Separately, in auto mode the classifier reviews the code as a whole rather than each call inside it, and Anthropic's own documentation warns that the classifier does not see values computed at run time. A script that searches first and then acts on a path the search returned is, in auto mode, reviewed before that path exists. Teams that rely on auto mode for unattended work should read that section of the docs to the end before they switch this on.

Version 2.1.325, one day after launch, fixed several Code Mode bugs, including scripts that ignored a worktree and a case where Claude began working on its own after a session was resumed, as reported from the changelog. For an experimental feature in its first week, that is a reason to treat the early enthusiasm, and its borrowed statistic, as early signals rather than results.

Sources

  1. [1] Anthropic, Claude Code documentation, "Run tool calls from code with Code Mode" (updated November 13, 2026)Read source
  2. [2] Anthropic, "Introducing advanced tool use" (November 2025), origin of the 43,588 to 27,297 token figure for Programmatic Tool CallingRead source
  3. [3] i10X News, "Claude Code Code Mode: JavaScript tool calls" (changelog and documentation relay)Read source
  4. [4] AgenticFeed, "Claude Code Code Mode: JavaScript tool calls"Read source
  5. [5] Liam Godoy, "Claude Code's New Code Mode Is Real. The Viral '37% Fewer Tokens' Number Isn't" (Medium, November 2026), documenting the 37% figure's circulation with Code ModeRead source
  6. [6] SecurityOnline, "Claude Code Mode Automates JavaScript Tool Chaining" (November 16, 2026)Read source