I spent $514 on Claude Code in 30 days. Here's what I learned.

Real data from 50 sessions. The good, the bad, and the expensive.

The setup

For the past month, I've been tracking every Claude Code and OpenCode session with claudestat — an open-source tool I built to monitor token usage, costs, and patterns across both CLIs.

This isn't a hypothetical study. These are my real numbers from real work on real projects.

The numbers

30-day period: April 12, 2026 → May 12, 2026

That's about $18/day on average. For context: I'm on Max 5, so this isn't cheap.

The most expensive session

April 26, 2026 — I let Claude run on a refactoring project for hours without checking:

• Cost: $32.94
• Loops detected: 31
• Efficiency score: 35/100
• What happened: Claude went in circles editing files, each edit triggering more edits. I wasn't watching.

That's $33 in one session. More than my daily coffee budget for a month.

Key lesson: Loops are expensive. 31 loops × average tool call cost = $33 down the drain.

Now claudestat sends a notification when a session closes — cost, tokens, loops detected, context % — so I always know what just happened.

The project breakdown

Who consumed the most?

63% of my spending went to one side project. It's not my main job — it's my hobby project that's burning my quota.

This is the kind of insight you only get from tracking.

The loop problem

375 loops detected across 50 sessions.

That's ~7.5 loops per session on average. But it's skewed:

• 12 sessions had more than 10 loops
• 9 sessions had efficiency below 50%

Most loops look like:

Read → Edit → Edit → Edit → Read → Edit → Edit → Edit

Claude tries something, it doesn't work, tries slightly different, repeats. This is where most of my waste came from.

Context limits hit me harder than quota

The $33 session wasn't killed by quota — it was killed by context. At 95% context window, Claude starts degrading. I had no warning.

Now claudestat sends alerts at 50%, 75%, 90%, and 100% context — and the same thresholds apply to your 5h quota cycle. You get warned before hitting the wall, not after.

claudestat status

Quota 5h   3.2h / 5h (64%)  |  reset in 1h 48m
Context    91,200 / 200,000 (45%)
Burn rate  1,240 tok/min  →  ~73min remaining
Plan       MAX5

What I learned

1. Loops are the silent killer

I wasn't aware of loops until I saw the data. Now I watch for:

• Same tool called 3+ times in a row
• Context creeping up while output stays flat
• "Let me try..." messages from Claude

2. One project dominates

63% on claudetrace — I would have guessed 40/60 at best. Tracking revealed the truth.

3. Efficiency varies wildly

• Best session: 100/100 (perfect)
• Worst session: 35/100 (that $32.94 day)

Understanding this helps me spot when I'm about to have a bad session.

4. Edit is my most expensive tool (not Bash)

Running claudestat top showed Edit at 21% of total cost, Bash at 20%. They're almost equal — but Edit surprised me. Large file rewrites are expensive.

5. Claude can now ask itself how it's doing

Since I added the MCP server, Claude can query its own stats mid-session:

"What's my current quota status?"
→ 64% used, 1h 48m until reset, burn rate 1,240 tok/min

"What are my top 5 tools by cost this week?"
→ Edit $146 · Bash $140 · Read $126 · Grep $39 · ToolSearch $21

This changed how I work. Claude now knows when to compact, when to stop, when to wrap up.

The weekly view

claudestat weekly

📊 Week of May 8 – May 13
💰  $198.38  ·  40 sessions  ·  114 loops
📈  Efficiency  91/100
💾  Cache hit   100%
⚡  Tip: 114 loops — consider /compact earlier

One number per week. Am I improving or not?

How to track your own data

npm install -g @statforge/claudestat
claudestat install
claudestat start
# use Claude Code or OpenCode normally
claudestat weekly       # weekly summary
claudestat top          # most expensive tools
claudestat status       # quota + context right now

What's next

Now that I have this data:

1. Alerts at 50/75/90/100% — context and quota, no more surprises
2. Session-close summary — always know what just happened
3. Check top tools weekly — spot expensive habits
4. Watch for loops — stop them early
5. Claude knows its own stats — MCP server installed, it manages itself better

The black box is open. You have the data.

Try it

npm install -g @statforge/claudestat

Works with Claude Code and OpenCode. No cloud, no accounts — everything stays local.

Questions? Drop them in the comments.

Data from my actual sessions. Not estimated. Not hypothetical.