The One Command That Halved My Anthropic Bill Overnight

Last month my Anthropic bill was $312. I use Claude Code for 6-8 hours daily across multiple projects. After adding a single line to my shell config, this month's projected bill is $94. Same usage patterns. Same quality of output. Same number of sessions.

The difference: I stopped sending redundant tokens to the API. That is it. No workflow changes. No prompting tricks. No switching to a cheaper model.

Here is the full breakdown of what happened and how you can do the same thing in 60 seconds.

The Setup (Literally 60 Seconds)

pip install copium-ai
copium wrap claude

That is it. Two commands. Now every Claude Code request routes through a local compression proxy before hitting the API. My prompts get 40-80% smaller. Same answers come back.

Where My Tokens Were Going

I ran copium stats --period month after the first week and saw the breakdown:

Most of my wasted tokens came from duplicate file reads (180K → 12K, 93% savings), JSON tool outputs (320K → 64K, 80%), build logs (95K → 14K, 85%), search results (150K → 30K, 80%), conversation history (200K → 140K, 30%), and tool schemas (45K → 8K, 82%). Overall, my daily input dropped from 990K tokens to just 268K, a 73% reduction.

The Cost Breakdown

Anthropic Claude Sonnet pricing:

Input: $3 per million tokens
Output: $15 per million tokens (unchanged by compression)
Cached input: $0.30 per million tokens (90% discount)

My savings come from two sources:

Fewer input tokens (compression)

More cache hits (prefix stabilization)

Daily input tokens dropped from 990K to 268K, while my cache hit rate increased from 12% to 48%. That reduced my effective input cost from $2.90/day to just $0.62/day. Output costs stayed the same at $7.50/day, bringing my total daily cost down from $10.40 to $8.12.

Wait, that is only $68/month savings on raw math. Where does the $200 come from?

The bigger savings: I stay in sessions longer without hitting compaction. Before compression, long sessions hit compaction at 35 turns, forcing context loss and repeated work. Now sessions last 55+ turns productively. Fewer repeated file reads, fewer redundant tool calls, fewer wasted output tokens on re-doing work.

Does Quality Actually Stay the Same?

I was skeptical too. Here is what I measured over 4 weeks:

Code that compiles first try: 78% (before) vs 76% (after) = within noise
Tests passing on first run: 62% vs 60% = within noise
"Agent forgot something" incidents: 4.2/week (before) vs 1.1/week (after) = BETTER

The last metric surprised me. Compression actually IMPROVED context management because the agent's context window was not overflowing with garbage.

What If I Have a Copilot Subscription?

Subscription users do not pay per token directly, but you still benefit:

Longer productive sessions (context does not fill up)
Fewer "I need to start a new chat" moments
Better quality in long sessions

Copium

Copium (http://github.com/iKislay/copium) is open source (Apache 2.0) and runs entirely locally. Your code never leaves your machine. It adds about 50ms of latency per request, which is invisible compared to the 2-30 second LLM response time.

The key features that matter for cost savings: