Home › Guides › Claude Code usage tracker

Is Claude Code nerfed? What 48 days of my own usage logs actually show

In 48 days, Claude Code went through $935 worth of tokens - on a flat monthly subscription. Twice in September I hit my session limit right in the middle of work, and everywhere I looked people had a theory: limits got cut, the new model thinks less, Anthropic quietly made it dumber. Nobody showed actual numbers. Then I realised I have them.

Where the data lives

Claude Code keeps a log of everything it does, on your own computer: a hidden folder in your home directory, ~/.claude/projects, one file per session. Every reply from the model comes with a usage block - fresh input, cache writes, cache reads, output, even the thinking tokens - plus which model it was and the effort level. So I built a small tool that reads all of it. It is one Python file with no packages, and it only reads files: nothing leaves your machine. 

python3 skill/usage-tracker/usage.py --redact

--redact swaps project names for placeholders, so you can share a screenshot. It prints a summary and writes an HTML dashboard: cost per day, the days you hit a limit in red, cost per model and per project.

The usage dashboard: 934.90 dollars API-equivalent cost, 584.31 per 30 days, 7,655 model calls, 2 limit hits, 37.93 used before a session limit, and a bar chart of cost per day with two red limit days.
7,655 calls over 48 days. The red bars are the days I hit a limit.

These dollars are not my bill. I pay a flat monthly price. They are what the same tokens would cost on the API at Anthropic's published prices - the best ruler there is, because Anthropic does not publish exactly how subscription limits are counted.

Where the money goes

This was the surprise. Only 13% goes to what Claude actually writes. 63% is Claude re-reading the conversation: every time it takes one step, it reads the whole session again from the cache.

Where the money goes: re-reading context (cache reads) 587.64 dollars, 63 percent; writing context (cache writes) 227.48, 24 percent; output plus thinking 119.71, 13 percent; fresh input 0.07, 0 percent.
Re-reading context $587.64 (63%), writing it to cache $227.48 (24%), output and thinking $119.71 (13%). Calls with more than 200K tokens of context account for 81% of the cost.

So, is it nerfed?

Week by week, the output per call is basically flat: around 400 to 500 tokens, every week. Claude is not writing less. What changed is the context. Two weeks before the video, my typical call carried 127,000 tokens; that week, 518,000. Four times more.

Week by week table: context per call rising from 127K to 518K tokens while output per call stays around 400 to 500 tokens, with the effort level and model per week.
Context per call (highlighted) grew from 127K to 518K; output per call barely moved.

The logs show one more thing: Opus 5 ran on extra-high effort, Opus 5.5 runs on medium. Per call, the newer model is actually cheaper.

The limit hits

The logs record every time you hit a limit and when it resets, so the tool adds up what you used in that five-hour window. My first hit came after $41 of API-equivalent usage, the second after $35. But the night before the video I had a $65 block and no limit at all. I cannot explain that from the logs - maybe my phone or the website counted too in September, since they share the same limit but are not in these files. So I will not pretend I know the exact limit.

What actually helps

That is where my limits really went. From my own data, Claude Code is not nerfed: the model writes about the same as before, and my sessions got four times longer.

Or just ask Claude Code

The tracker is also a Claude Code skill. Copy it into ~/.claude/skills/ and ask "am I being nerfed?" or "where did my usage go this week?" - it runs the tool, reads the numbers, and answers with them.

On Patreon · $9 one-off

Claude Code usage tracker - the script and the skill

Your own numbers in about ten seconds, from logs that are already on your computer.

  • usage.py - one Python file, standard library only: terminal summary plus an HTML dashboard.
  • The Claude Code skill - ask "am I being nerfed?" and get an answer with your numbers.
  • Every view from this page - cost per day, model and project, where the money goes, limit hits, biggest 5-hour blocks, week by week.
  • --redact to hide project names before you share a screenshot.
Get it on Patreon →

Patreon members ($5/month) get it with all my other tools, or buy just this one for $9.

Honest limitations

Watch the video

The dashboard, where the money goes, the week-by-week check, and asking Claude Code directly.

FAQ

How do I check my Claude Code usage?

Claude Code logs every model call in ~/.claude/projects on your computer. The tracker reads those logs and shows cost per day, per model and per project, where the tokens go, and every time you hit a limit.

Is Claude Code nerfed?

Not in my data: over 48 days, output per call stayed flat at about 400-500 tokens a week. What grew was my context per call, from 127K to 518K tokens - longer sessions, not a weaker model.

Why do I hit Claude Code limits so fast?

In my logs, 63% of the cost was Claude re-reading the conversation, and 81% came from calls with more than 200K tokens of context. Fresh sessions per task, /compact and a short CLAUDE.md help most.

Does the tracker send my data anywhere?

No. It is one Python file that only reads the local log files and writes a local HTML report.

Related

Free Wispr Flow Alternative → Hold a key, talk, get clean text - fully offline on Mac, Windows and Linux. Claude Code for free with a local model → Ollama, the three traps, and a benchmark against the real Claude. All guides → Zabbix, Grafana, AI and the tools from my Patreon, written from the videos.

Want to see where your team's AI spend goes?

Measuring usage from the data you already have, instead of guessing, is most of what I do at KorFlux - for AI tools, infrastructure and monitoring alike.