AI

What I use every day, and how much.

Models

Token share, last 30 days

Claude Opus 5.5My default for almost everything.93.9%

Claude Fable 5.1My second most used model this month.4.6%

GPT-6.1 SolMost of my Codex tokens.0.9%

Harnesses

Token share, last 30 days

T3 CodeWhere I run my agents. Its tokens count under Claude Code and Codex.

Claude CodeAnthropic's coding agent.98.7%

CodexOpenAI's coding agent.1.3%

tstack

Private for now

My own Michelin kitchen, as poteto calls it: the skills and rules that make agents write code, review PRs and ship features the way I would. When I correct an agent, the fix becomes a rule, a check or a skill, so I don't have to say it twice.

Inspired by poteto's pstack.

code-like-thalles
review-pr
babysit-pr
correct
improve-skills

What I used before

How I got to the setup above, in order.

GitHub Copilot
Where everyone started. It was revolutionary, and then it barely improved. GitHub fumbled the lead.
Supermaven
Like Copilot, but a lot faster and with much better predictions.
Cursor
Where I really started using AI, and tokens. Quick edit first. I ignored the sidebar at the start, then it became my default.
Claude Code
I flirted with it, but I prefer tools that wrap it. Opus also scored better inside Cursor than inside Claude Code.
OpenCode
My main for a while, because I like jumping between models.
Cursor, again
Back for their agents in the browser.
T3 Code, the first time
I tried it while looking for something that worked with a remote dev machine, which I was already doing by February. It wasn't there yet.
Maestri
I loved it and daily drove it for about two months. In the end it was too complicated, and I got lost in my own workflows.
T3 Code, now
I came back for the sidebar, which is amazing. It's open source, runs a bunch of harnesses, performs better, and remote dev on it is great. It's what I use every day.

How much AI I use

Most of my code is written with agents now. This is how many tokens that takes.

Tokens per daySep 29 to Oct 28

Tokens, all time: 62.5B
Output, 30 days: 107.7M
Sessions, all time: 3,205
Top model, all time: Claude Opus 5.5

Cost at API list price would be $52,526. I pay a subscription, not that. Tracked with tokenmaxxing.sh.