Start here
I design AI agents, software and automations, and I publish what I test. This page has the weekly newsletter, the list-cost table and four recent articles.
The weekly newsletter
One email a week with what I test and the list-cost table already measured: list USD, not the amount billed. Three models, one task, October 2026.
List cost of one task
Read the full measurementThe task was to build a 'roadster' game. The prompt was the same, with one run per model.
- $1.14Astra · 5 min 50 s
- $1.21Kimi · 34 min 42 s
- $3.10Opus · 17 min 4 s
Task: the same roadster game, one run per model, October 24, 2026. List cost, not the amount billed.
| Astra | 5 min 50 s | $1.14 |
|---|---|---|
| Kimi | 34 min 42 s | $1.21 |
| Opus | 17 min 4 s | $3.10 |
Other tests on the blog used different tasks and give different numbers.
List USD, not the amount billed. Same prompt, one run per model, October 2026.
Four articles to continue with
‘--dangerously-skip-permissions still won’t let rm -rf through’: what’s actually proven, and how to protect your machine
A Claude Code engineer announced CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, an env var that turns off the safety net against rm -rf ~. It first appeared in a post and in npm 2.1.311; the official docs now describe it. Checklist for --dangerously-skip-permissions.
GPT-6.1 Terra: does it really cost $2 per million tokens?
GPT-6.1 Terra costs $1 input and $5 output per million tokens. The $2/$10 going around is Fast mode. What it ships with, what OpenAI claims and when it is worth using.
How I Tripled What I Earn as a Developer Using AI
My real AI workflow as a developer: the problem of paying for AI that doesn't pay off, 3 rules for picking tool, model and effort level, the mistake that burns the most money, and one bounded cost test.
They said phones weren't listening.
Today it's a Samsung TV. Before that it was Apple, Google and Amazon. Here's what's proven, what's alleged, and what's still a rumor, plus the settings to change tonight.
The weekly newsletter
One email a week with what I test and the list-cost table already measured: list USD, not the amount billed. Three models, one task, October 2026.