Been all-in on Codex CLI with GPT 5 high for the past few days. It's really surprising to me how much better the same model is in CLI (vs. IDE/web UI).
1
1
2
I suspect it's because the IDEs expose different interfaces to interact with code & IDE tooling, which the models aren't as good with. Less training data, polluting context, etc. The model can just cook in the CLI. (This is just conjecture. Would love to know if that's wrong. )
1
1
Does make me feel like "developer experience" will mean something very different going forward. Models DGAF about pretty, they GAF about data/efficiency/etc.

Nov 9, 2025 · 12:45 AM UTC

1
1
3
Replying to @criccomini
I suppose most will lean into using model providers directly that expose the models without much system prompt so we can customize our workflow with our context (skills mcp etc.)
1
Load more