semisol on Nostr: I’d recommend trying OpenAI models. The flexibility with your own harness is worth ...

I’d recommend trying OpenAI models. The flexibility with your own harness is worth it.

5.3 Codex was the first time they had a well-polished coding model, imo. But it always needed high/xhigh.

5.4 was much better on a lot of fronts:
- Better usability in terms of personality.
- Much better technical capability than 5.3 on hard tasks
- Can do a lot on high/medium, xhigh not needed
and 5.4 mini can replace 5.3 Codex easily.

5.5 fixed a lot of the “early stopping”, and had better depth and technical capability.
The main advancement was token efficiency and medium/low(!) working for 90% of tasks.

5.6 is much more in depth with tasks, works well on low reasoning levels, and seems to have significantly better capabilities.

quoting
nevent1q…ddjm

5.6 feels much better than 5.5.

It goes into further depth on a task, and has very little “early stopping”.