← All reviews
Analysis · November 15, 2026 · 4 min read

Claude Code's Code Mode, tested: about 22% cheaper on two of three tasks

Claude Code's experimental Code Mode cut the cost of a feature-and-tests task by 23% and a broken-link check by 22% in our runs. It runs only in an interactive terminal session, so try it there on multi-step work.

The 30-second version

Code Mode cut the cost of two of our three tasks by about a fifth at list prices, and made none of them slower.

  • Feature with tests: $0.35 a run instead of $0.45, done in about 60 seconds instead of 80. All six runs passed the tests.
  • Broken-link check: $0.16 instead of $0.20, in 18 seconds instead of 26.
  • 20-file edit: about the same, $0.16 against $0.17.
  • Why. The first request was about 16,000 tokens smaller, and the feature task needed 6 to 8 requests instead of 10 to 13.

Three tasks, three runs each

Each run started in a fresh copy of a public repository with the same prompt. Code Mode on and off ran side by side, on Opus 5.5:

TaskCode Mode onCode Mode off
Add a --max-per-source option with tests$0.34–0.36, 57–63 s, 6–8 requests$0.43–0.47, 71–97 s, 10–13 requests
List links to local files that don’t exist, in 134 Markdown files$0.14–0.18, 17–22 s, 3–4 requests$0.18–0.22, 22–30 s, 4–5 requests
Add version: 1.0.0 to 20 SKILL.md files$0.16–0.17, 13 s, 3 requests$0.16–0.17, 13–15 s, 3–4 requests

The test suite passed after every feature run, every link check listed the 8 broken links we had planted, and every edit run changed exactly the 20 lines asked for.

Where the saving came from

Code Mode’s sessions started smaller. Claude got 3 tool definitions, REPL, Poll and advisor, instead of 16; the first request was about 22,000 tokens instead of 38,000. Every later request sends that base again, mostly as cache reads.

On the feature task Code Mode also cut round trips: Claude made 6 to 8 requests instead of 10 to 13, and some of its code made several edits and ran the tests in one go. Cache reads fell from 512,000–678,000 tokens to 211,000–274,000, and output from about 6,000 tokens to 4,800.

On the 20-file edit, smaller cache reads were offset by slightly larger cache writes. Without Code Mode, Claude wrote scripts for the link check and the edit and ran them with Bash; with it, Claude did the same work in JavaScript.

Before you turn it on

Start Claude Code 2.1.323 or later with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 and confirm once. Anthropic’s documentation lists the limits:

  • Where it runs. Only an interactive claude session in a terminal, on a subscription or Console account. Not -p, the Agent SDK, the VS Code panel, the Desktop app or cloud sessions.
  • What stops it. DISABLE_TELEMETRY, OpenTelemetry export, an API key helper, agent teams, or a gateway or cloud provider.
  • What you give up. Remote Control, Claude in Chrome and computer use. A Code Mode session resumes only with Code Mode on.
  • Permissions. Your rules and hooks still check each tool call, but auto mode’s classifier reviews the code once, not each call.

How we measured

  • Version and model. Claude Code 2.1.325, installed separately, with Opus 5.5 at its default effort, medium, on 15 November 2026. Permissions were skipped and user settings weren’t loaded.
  • Tasks. The feature task used the public lastdays-skill repository at commit 0ff966d. The other two used a copy of anthropics/skills at 1780c5e, saved without its git history, with four linked files deleted to plant 8 broken links.
  • Cost, tokens and requests. Claude Code’s own session totals at list prices, and the requests counted in each session’s transcript. Time runs from the prompt to Claude’s last reply.

What we didn’t test

  • Plan limits. Other sessions shared the account, so we didn’t read the 5-hour meter.
  • Other models, efforts and long sessions. Only Opus 5.5 at medium, on tasks that took under two minutes.
  • Your own setup. With MCP servers, plugins or permission prompts, both the tool list and the number of requests can differ.

Companion reading

Sources

  1. Anthropic — Run tool calls from code with Code Mode. Retrieved 15 November 2026.
  2. Anthropic — Claude Code changelog, 2.1.323 and 2.1.325 entries.
  3. Anthropic — Pricing. Retrieved 15 November 2026.
  4. GitHub — anthropics/skills and nvwalj/lastdays-skill, the repositories we ran the tasks on.

FAQ

Does Claude Code’s Code Mode save tokens? In our runs on Opus 5.5, yes: it used fewer tokens on all three tasks, and at API prices it cost 23% less on a feature-and-tests task and 22% less on a broken-link check.

How do I turn on Code Mode in Claude Code? Start Claude Code 2.1.323 or later in a terminal with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1, or put that variable in the env block of ~/.claude/settings.json, then confirm the prompt once.

Does Code Mode work with claude -p or VS Code? No. Anthropic’s documentation says Code Mode works only in an interactive claude session in a terminal, not with -p, the Agent SDK, the VS Code panel, the Desktop app or cloud sessions.

Why won’t Code Mode start? The documentation lists DISABLE_TELEMETRY or another variable that turns off feature-flag fetching, OpenTelemetry export, an API key helper, agent teams, a gateway or cloud provider, and versions older than 2.1.323.

Was this helpful?

Related reading


Reviews independently produced · Editorial policy

Read more reviews →