Claude Code Token Costs: The Hidden Tax and How rtk Cuts It by 80%

Every bash command Claude Code runs has a hidden token cost. git status outputs ~3,000 tokens. cargo test can output 25,000 tokens. A typical medium project development cycle burns through 118,000 tokens just on command output — and with a 1M context window, that cost compounds silently until your daily quota vanishes.

The Problem: Token Billing is a Black Box

A user tracked GitHub issue #16157 and found a striking case: 92% of tokens in a session came from cache reads, actual output tokens were near zero, yet the API charged $1.50 — which counted as $65 of usage against their quota.

Why This Happens

The 1M context window is an amplifier. Key pain points:

Quick Workarounds

• /compact — Compress context before it gets too large; don’t wake old sessions, start fresh
• /cost or /stats — Monitor token consumption in real-time
• Don’t revive stale sessions — Opening yesterday’s conversation reloads the entire history

The Solution: rtk — CLI Proxy That Saves 80% of Tokens

rtk (Rust Token Killer) is a single Rust binary that sits between Claude Code and your shell commands. It intercepts command output, compresses it, and sends Claude only the essential information.

Before rtk:           After rtk:
git status → 3000 tokens    → 150 tokens (95% saved)
cargo test → 25000 tokens   → 2500 tokens (90% saved)
ls -la     → 800 tokens     → 150 tokens (82% saved)
─────────────────────────────────────────────────
Total: 118,000 tokens  →  23,900 tokens (80% saved)

How rtk Works: 4-Step Compression

Raw command output
       │
       ▼
┌─────────────────────┐
│ 1. Smart Filtering   │  Remove comments, whitespace, boilerplate
├─────────────────────┤
│ 2. Grouping          │  Cluster similar items (e.g., files by directory)
├─────────────────────┤
│ 3. Truncation        │  Keep useful info, cut the rest
├─────────────────────┤
│ 4. Deduplication     │  Fold repeated log lines into one + count
└─────────────────────┘
       │
       ▼
Compressed output → Claude

Setup: One Command

# Install on macOS
brew install rtk

# Then initialize globally
rtk init --global

After this, git, cargo/npm test, docker, kubectl, eslint/ruff/pylint, jest/vitest/playwright, pip/pnpm — all common dev commands are automatically intercepted. Claude sees only compressed versions, completely transparently.

Real Savings: Why It Matters Beyond Cost

Token savings aren’t just about money:

• More context space — Compressed outputs mean more room for code history, more context for better Agent decisions
• Longer sessions — Same conversation can stay productive longer before hitting limits
• Better agent performance — Less noise in context = more focused, accurate responses

Limitations

Workaround for built-in tools: Use shell equivalents explicitly — cat/head/tail instead of Read, rg/grep instead of Grep — when you want rtk compression. Or explicitly call rtk read.

Supported Platforms

Claude Code, OpenCode, Gemini CLI — install is one command, config lives at ~/.config/rtk/config.toml.

The Other Side: Cut Output Tokens with claude-token-efficient

rtk cuts input tokens (command output → Claude). But Claude also wastes tokens on output — filler phrases, flattery, restating your question, unsolicited suggestions. claude-token-efficient is a CLAUDE.md snippet that eliminates this waste.

The Problem: Default Claude Output Is Bloated

By default, Claude:

All of this wastes tokens. None of it adds value.

The Fix: One File in Your Project

Drop the snippet into your CLAUDE.md. Claude immediately becomes more concise — fewer tokens per response, faster output, cleaner parsing.

# Add to any project
curl -O https://raw.githubusercontent.com/drona23/claude-token-efficient/main/CLAUDE.md
# Or just add the rules to your existing CLAUDE.md

When to Use vs When Not To

Combined Strategy: rtk + claude-token-efficient

Input tokens:   rtk compresses command output      → 80% saved
Output tokens:  claude-token-efficient cuts filler  → 30-50% saved
─────────────────────────────────────────────────────────────
Combined:       Both sides optimized                → significant cost + quota savings

Use both together for automated pipelines. For interactive work, rtk alone is usually enough — you probably want Claude’s explanations when you’re learning.

How LearnAI Team Could Use This

• Teach token cost literacy as part of AI engineering practice.
• Use rtk examples to show how tool output affects agent performance and budget.
• Build classroom exercises where students compare raw versus compressed command output.

Real-World Use Cases

• Reducing token burn during test-heavy coding sessions.
• Keeping long Claude Code sessions usable on medium or large repositories.
• Teaching teams to monitor and compress noisy CLI output before it enters agent context.