Max Tokens, Max Thinking
Just bought ChatGPT Pro at $200/month — felt like I had to max it out to get my money’s worth. Plus everyone knows gpt-5.4 xhigh is the strongest coding model. So I cranked everything to max. More thinking must be better, right?
What Broke
After a stretch at high intensity — 500M tokens a day, 2–4 windows, each with 0–3 subagents — the cracks showed.
Slow. xhigh is genuinely stronger, but it’s also genuinely slow. For everyday coding, the idea has already gone cold while the model is still deep in thought.
Premature optimization. Most obvious in glue code: it invents compatibility for legacy logic you don’t need, and plans for contingencies that aren’t on the roadmap. Progress drags, code doesn’t get better.
Wait forever, get a worse result.
Thinking via OMO
I’d brushed past oh-my-opencode before without getting why model orchestration matters.
The point of subagents isn’t more thinking — it’s division of labor. Right agent for the right job.
OMO lets you assign different models to different roles:
| Role | Responsibility | Model |
|---|---|---|
| Sisyphus | overall coordination, task breakdown | gpt-5.4 high |
| Hephaestus | hands-on coding | gpt-5.4 high |
| Oracle | architecture decisions | gpt-5.4 high |
| Explore | fast codebase search | gpt-5.4-mini-fast medium |
| Prometheus + Metis | upfront planning | gpt-5.4-mini-fast high |
| Momus | rigorous review, nitpicking | gpt-5.4 xhigh |
| visual-engineering | frontend visuals | gemini-3.1-flash-lite-preview high |
| Librarian | knowledge retrieval | gpt-5.4-mini-fast high |
| Atlas | deploy & ops | gpt-5.4-mini-fast high |
Simple logic: reserve xhigh only where deep thinking matters.
- Momus on xhigh: review demands depth.
- ultrabrain on xhigh: architecture planning demands depth.
- Execution, search, everyday tasks drop to high or medium.
Speed returns, flow stays intact. And where depth is needed, it’s still deep.