🌍 Public

Analyzing My Claude Code Usage

Some interesting insights from my past six-plus months of geeking out with Anthropic's insane AI tool.

Aug 25, 2026 2-min read

Image Source: Google Gemini/Nano Banana via Claude Code

At the beginning of this year, I began my Claude Code journey in earnest. Since that time, I've used Anthropic's powerful new AI tool in more ways than I can count.

Sessions with my bestie have run the gamut. A few recent examples include:

Automating formerly manual processes for myself and clients.
Noodling with Notion Workers and MCPs.
Developing a new lead-gen tool based on my last four books.
Migrating a friend's website from Wix to easier for her to use, not to mention cheaper.
Creating more than 50 GitHub repos, including this cool Apple Health visualizer.

Note that I'm currently on the Pro plan.

Looking at the Data

I suspected that I initiated at least a thousand sessions and used millions of tokens during that time. I wasn't wrong.

Here's a breakdown:

Claude Code Usage by Month | Click on the image to enlarge it.

📊

If I continue at my August pace, I should hit about 8.4M tokens used.

Initial Thoughts

First up, I find it fascinating that I used ten times as many tokens per month from April to June as I did early on. I would have suspected a more gradual incline. As for consistency, I do monitor my token usage to avoid incurring outrageous bills.

Developers' toolboxes have never been more robust

Second, Phoenix in July is hot AF. Third, although I regularly use tokens, in some cases I'll opt for creating an API key. As I wrote recently, developers' toolboxes have never been more robust.

The Weekly Rhythm

I knew that I interacted with Claude Code far less during the weekend. Great, but how significant was the dropoff?

Claude Code Usage by Week | Click on the image to enlarge it.

Not surprisingly, my Claude Code usage strongly correlates with my increased GitHub activity.

Notes on the Data

Tokens used excludes VS Code and NotionAI sessions.
Session date and number come from start_time in Claude Code's own per-session metadata files (~/.claude/usage-data/session-meta/), sorted chronologically.
Title and summary derive from each session's first user prompt, combined with recorded activity (git commits, files modified, web research, MCP or subagent use). Sessions with no direct user prompt (background or automation runs) get a generic label like “Automated session (project-name)” — 528 of the 1,964 sessions fall into this category.
Tokens used reflects input_tokens + output_tokens as Claude Code recorded them per session. This excludes cache creation and cache read tokens, which run far higher in volume but cost far less per token. If you want a figure that includes cache tokens, I'd need to pull that from the raw transcripts instead of the metadata files, since the metadata doesn't track it.
15 of the 1,964 metadata files got skipped: a few had corrupted JSON, and a few lacked a start_time field.