So I'm building an app for vibe coders with cloudflare codemode and the cost saving for tokens is literally a game changer.
Same task cost:
> $0.018 with code mode
> $0.26 without it
If you don't wanna ship AI slop, then planning should be done in very detailed mode.
Here's I created 7 plans from start to finish.
Each plan has 4-5 phases and then each phases split into M0, M1, M2 etc...
All these 7 plans created by claude code which understand my whole project from start to finish.
And then I use the same memories collected by claude code and put it in codex (gpt 5.5) to for additional review and findings to make the plan more robust.
Finally then I use 4-5 agents and ask each agent to build M.x of each phase and then human review of each phase.
This reduce the AI slopiness, boilerplate code it produces.
Every single time gpt 5.5 solve the hard problems in just one pass while opus struggling and ask a lot of questions and even hallucinate a lot.
But still I'm subscribed to $100 opus plan and $20 codex plan - just for only reason which is UI.
Hope gpt 5.6 solve the frontend UI.