# Is Haiku 5.5 good enough to be your main coding agent, not just the cheap one? I ran a small controlled test — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/8be9dfe1ac649beb

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-10-08 21:16 ingested 2026-10-09 02:39

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1x132yy/is_haiku_55_good_enough_to_be_your_main_coding/)

## retained text



```
I tested Claude Haiku 5.5 at low, medium and high effort against Sonnet 5.5 (medium), GPT-6 Luna (high) and GPT-6.1 Sol (medium). The tasks came from a real mid-sized, mixed-language codebase (Rust core plus a legacy scripting layer).

The tasks:  
two Rust bug fixes from vague bug reports  
a code review with 9 planted bugs and 2 decoys  
two Rust implementations from tricky specs  
a "judgement" task, where the bug report blamed the wrong thing and asked for a fix that would make it worse, and one requirement was deliberately ambiguous  
an agentic task in the real repo: move some logic across a language boundary following the repo's unwritten conventions

I wrote the answer keys and hidden tests before any model ran.

The surprise: on the well-specified tasks every model scored 100%. Haiku-low missed one compile error in the review, and Luna asserted its own reading of the ambiguous requirement where the others flagged it, but that was it. Self-contained coding tasks just don't separate current models any more.

The agentic task is where they separated. Haiku at medium did as well as Sonnet: correct code, all tests passing, the repo conventions followed, an honest note on what it couldn't verify, and it raised a real design trade-off on its own. It made 29 API requests and cost about $0.08, against Sonnet's 30 requests and about $1.68 at list prices.  
The other effort levels did worse:

**Low** got working code, but skipped verification, missed an error-handling fallback, and invented a reason for not testing ("the build stalled for over an hour" in a run that took 15 minutes).

**High** was correct, but burned 385 requests and about $1.03 for no quality gain.

My takeaway: Haiku 5.5 at medium is viable as a general daily coding agent, not only for low-stakes work, at roughly a twentieth of Sonnet's cost. Avoid low effort for agent work, high effort buys nothing, and I'd still use a bigger model for high-stakes release work.
```



document fetched 2026-10-09 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1x132yy/is_haiku_55_good_enough_to_be_your_main_coding
