Claude Code's concise output mode drops token usage by 40% in my

Been running the new concise style for about a week now across a few different projects — React frontend, Python backend, some infrastructure scripts. The difference is noticeable immediately. Standard mode tends to overexplain every file change, repeat context I already gave it, and pad responses with "here's what I'm doing" preamble. Concise strips all that.

What actually changes

The model still reasons through the same steps internally — you can see it in the thinking traces if you enable verbose logging — but the final output sent back to you is compressed. Function signatures without docstring repetition. Diff hunks without the surrounding "I'm now editing X file" commentary. Error explanations cut to the actionable fix.

# Before (standard)
I'll update the authentication middleware to handle the new token format. 
Let me first examine the current implementation...

# After (concise)
Updated auth middleware for new token format. Changed `validate_token()` 
signature in `middleware/auth.py:42` to accept `token_version` param.

Token savings add up fast. A typical 5-file refactor that used to burn 8-12k output tokens now lands around 4-6k. Over a full day of coding that's real money on the API bill.

Where it falls short

Two scenarios where I flip back to standard:

1. Debugging unfamiliar codebases — when I'm dropping into a repo I've never seen, the extra context in standard mode helps me understand why the model made certain choices. Concise assumes I already know the architecture.

2. Complex multi-step planning — if I ask "plan the migration from REST to GraphQL across these 12 services," I want the full reasoning trail. Concise gives me the task list but not the tradeoff analysis.

Config toggle

# .claude-code/config.json
{
  "outputStyle": "concise",
  "verboseLogging": false
}

Or per-session with /style concise in the CLI. I keep a shell alias cccon that launches with concise mode + my preferred model (sonnet-4) + the project context file preloaded.

One gotcha

The concise mode sometimes drops file paths in multi-file edits if the change is trivial (single-line fixes). Learned this when reviewing a PR — the model had updated three config files but only showed the diff for one. Now I run /diff after any batch edit to verify coverage.

Worth enabling if you're comfortable reading diffs and don't need hand-holding. Cuts noise without losing signal.