# Stop paying Opus prices to read ls output — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/0b5f04b074630562

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-31 12:12 ingested 2026-11-01 02:40

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xk8mzy/stop_paying_opus_prices_to_read_ls_output/)

## retained text



```
I pulled up a week of agent traces and sorted by cost and the most expensive calls were the model reading directory listings, grep results, test logs and file dumps.

The planning, writing code, deciding what's wrong was a small slice of the bill, the rest was the smartest model I could buy squinting at stack traces so it could summarize them for itself.

What fixed it:

\- Split the grunt work out like searching, reading files and summarizing logs go to a small model. It hands back a short summary and the big model never sees the raw dump

\- Don't switch models mid-conversation coz your cached context is the cheapest thing you have. Moving it to another model costs more than you save

\- Switch at task boundaries like a fresh subtask has no cache to lose, so that's the free place to downgrade

\- Watch tool calls as cheap models fumble them more and a quiet retry eats your savings

I run the small model side through Routerplus and any setup where you can point subagents at a cheaper model gets you most of the way.

What's the useless thing your agent is paying frontier prices for?
```



document fetched 2026-11-01 00:36 · canonical https://reddit.com/r/ClaudeCode/comments/1xk8mzy/stop_paying_opus_prices_to_read_ls_output
