After hearing about how gpt-5.4 is better at coding than Opus 4.6 from many people over the last few weeks, yesterday I started giving it a try.

I had begun a book graph project. It consists of graphing what books mention what books and vice-versa. Started it with Opus 4.6, now will carry on with gpt-5.4 at least for a few days.

Using codex over cli.

Very initial impressions: gpt feels faster, snappier.

I like the options we get before signing off on a command:

│ ❯ Yes (y) │
│ Yes, always approve this exact command for this session (a) │
│ Explain this command (x) │
│ Edit or give feedback (e) │
│ Switch approval mode (s) │
│ No, and keep going (n) │
│ No, and stop for now (esc)


First, you get introspection without interrupting your work flow. Second, the keys to hit are more ergonomic than Claude’s 1 / 2 / 3.

About the model itself, I haven’t felt the difference yet.

Small ideas not to forget

Ask LLM to cross-fertilize domains: “give me SEO ideas inspired by botanics”.

Brainstorm and do that at least 5 rounds deep, including going deep and wide in the iterations.

Ask LLM how it rates you as an agent operator. Ask it how you can do better.