Is Opus 4.8 Med really the overall best (smart+token optimal) for Code?
Hi! I’m not a programmer but a designer. I slowly learnt the tricks and got better at building. Now my GitHub is connected and I work in trees and have a push ritual. My Claude mds exist for each page. I have a budget for them too and recently I updated my release skill to optimize token usage during the final run before session ends.
Now my experience has been that Opus 4.8 on Medium does a fantastic job. I tried Sonnet 5 but it burnt a lot more tokens than Opus. I guess it’s new and still not baked in properly.
After multiple attempts, for Claude Code I am now permanently on Opus 4.8. Sonnet 4.8 messes up implementation sometimes even if a plan is drawn up. That leads to more rework and token burn. Opus runs clean. It has been doing this spinning up bots on the side and that is also a new thing that I’m not a fan of. I have to ask it to not do that.
My question to the community is this: am i doing it right? Am I misunderstanding use of haiku and sonnet? Is there a better way?
My main uses. On a max plan. Mac app.
- change and add modules on my existing portfolio site
- create fully functional small artifacts/utilities whose designs are half done in claude design or figma make
- apart from code, for normal chat: planning my sales calls, doing discovery, brainstorming demos, etc.
One last thing: chat is always the smartest for me when it comes to understanding context as opposed to cowork. Cowork can do good number crunching and all that but chat is the best when it comes to research and synthesis.
Please tell me where I’m going wrong and what I could improve.