About CodeTwig

Anthropic publishes benchmark numbers for every model, but they are hard to translate into everyday work. CodeTwig exists to show what actually happens when an ordinary person uses Claude Code on an ordinary task, with the cost, the diff and the pass or fail right next to it.

This isn't about saying Anthropic is wrong. It's about showing that every model is useful, and that you don't always need the most token-hungry one. With the right prompt or setup you can get more for less, and each recipe here is a small experiment that tests that.

We write down the task, run it in a fresh repository with Claude Code, keep the transcript, the diff and the cost, and then write up what happened, including the runs that didn't go the way we expected.

How a recipe is made

1. The recipe runs in a fresh repository, and the diff and cost are kept. Pass or fail is checked by tests Claude never sees.
2. We check the write-up against that evidence and the official docs before it goes live.
3. Only then is it marked verified, with the Claude Code version, model, date and cost. Until then it says "Not yet run".
4. When Claude Code or a model changes, we re-test. If a result changes, the badge says "Needs re-test" until it is run again.

Every recipe links to its run data: one row per run, each check, and the diff. If a number looks wrong, that is where to look, and we'd like to hear about it at victor@codetwig.dev.

Independent site. Not affiliated with or endorsed by Anthropic.