# I tested the "Haiku codes, Opus advises" setup everyone's sharing to save tokens. It was the most expensive of 5 setups. — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/dc76533ca91d3006

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-13 16:22 ingested 2026-10-14 02:42

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1x5167d/i_tested_the_haiku_codes_opus_advises_setup/)

## retained text



```
The setup going around this week: run Haiku as the main model and let it call Opus as an advisor when it gets stuck (claude --advisor opus). The idea is you pay Haiku prices for most of the work.

I ran it against 4 other Claude Code setups on the same real bug (from python-humanize). Git history stripped, no web access, graded by hidden tests the agent never saw. 3 runs each, medians, every run repriced from its token logs at Anthropic's API list prices.

- Sonnet alone: $0.15 per fix, \~1 min
- Haiku alone: $0.23, \~6 min
- Opus + Haiku subagent: $0.23, \~7 min
- Opus alone: $0.52, \~2 min
- Haiku + Opus advisor: $0.69, \~7 min

All 5 scored about the same (\~62/66 hidden tests). In the advisor setup, Opus was \~90% of the bill. The cheap model wasn't the problem, how often it called Opus was.

Caveats:
- one bug, 3 runs per setup, so treat it as a data point, not a verdict
- one advisor run got cut off before Opus answered, so that row is effectively 2 runs
- on easy bugs the ranking flips: Haiku alone was by far the cheapest (posting that next)

Has anyone measured how often their advisor actually fires? Curious if it pays off for you on harder or longer tasks.
```



document fetched 2026-10-14 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x5167d/i_tested_the_haiku_codes_opus_advises_setup
