Making my Codex usage go further

Even on the max $200 plan from OpenAI, I’m still scrambling to not run out of Codex tokens by the end of my usage period.

So I’ve started looking more closely at how I’m using those tokens and experimenting with ways to make them go further. Here are a couple things I’m trying right now:

The first is using ChatGPT Pro to do more of my planning and reviewing. It’s a better model for a lot of that work anyway, and it doesn’t eat into my Codex usage.

In practice, this means linking the GitHub repo inside a ChatGPT conversation, switching to the Pro model, and having it either do a planning session before I start coding or a review session after the work is done. I’m not having it actually write the code right now, although I may experiment with expanding into that later.

The second thing I’m trying is using more sub-agents running lower-reasoning models, with an orchestrator layer managing who gets what work.

Instead of having one expensive model chew through everything, I have one main thread farming work out to sub-threads or sub-agents based on the complexity of the task. Super simple work may get sent to Luna Max, mid-range work can go to Terra High, and the harder problems get funneled to Sol Medium.

The idea is pretty simple: don’t spend premium reasoning tokens on work that doesn’t require premium reasoning.

I’m still experimenting with all of this, but I’m becoming a lot more intentional about treating tokens like a resource that needs to be managed instead of something I can just burn through without thinking about it.

I’d be curious what other heavy Codex users are doing to stretch their usage further.