I reduced my Astra usage by 94% on a 23-hour build by making V4.1 Flash a native sub agent

[ethanplusai]

The first week Astra came out, the usage was incredible. I ran something like 500M tokens on the 20x plan and never got close to the weekly limit. Then that stopped. I started hitting the ceiling constantly and burned through four resets just to cover the next 500M tokens.

So I started trying to optimize for usage.

If you think about a long coding session, maybe 10% of it is actual judgment: architecture, interfaces, security calls, deciding if the result is right. The other 90% is read the repo, write the function, run the test, read the failure, fix it, run it again. Astra is worth every credit for the 10%. I was paying Astra prices for the 90%.

So my approach is to offload the "busy work" to lesser, yet still capable models.

First I reused orchestration skills I already had from other agents to push tasks to the smaller models. Maybe 10% savings, and it created new problems. Then I tried Sol, Luna, Opus, and Sonnet as the worker. Each had pros and cons, and none of them moved the number much.

Then I tried DeepSeek V4.1 Flash. The output was immediately better than the other cheap options, comparable to what I got from Opus 5, and it could hold a long task without wandering. That alone cut my Astra usage by about 60%.

But I realized the worker was never the real problem. Astra was still supervising: dispatching, checking in, re-reading, re-reviewing. A cheap worker saves you nothing if the expensive model is still parked in the loop watching it work.

So I rebuilt the orchestration so Astra gets out of the way. It plans once, dispatches once, waits, and reviews the finished patch in one batch. Flash owns the entire implementation, including testing and debugging its own failures.

That's when it clicked. A 7-hour build used about 2% of my weekly usage. A 5-hour build the old way had used over 28%.

I kept going on the same project for another 12+ hours to see if it held up on harder work. Full run results:

• 23 hours of active runtime

• 4 prompts from me for the main 18.5-hour stretch

• 94.2% less Astra input per 1K lines vs. doing it all on Astra

• $19.32 total on the DeepSeek side

It all runs inside Codex. Codex Router adds Flash as a native subagent, so there's no second CLI, and Astra reviews the actual diff rather than trusting the worker's summary. The installer does a dry run by default, backs up everything it touches, and has a clean undo.

If you want to check it out it's here on Github: https://github.com/ethanplusai/astra-flash-orchestrator

Please let me know if you have any feedback or find room for improvement.