← sources

I tested Haiku 5.5, Sonnet 5.5 and Opus 5.5 as subagents on 6 real tasks

source: reddit (reddit:ClaudeCode)published 2026-10-09 06:39ingested 2026-10-10 02:37

retained text

I took 6 tasks that Opus 5.5 had already done in my project and ran them again on Sonnet 5.5 and Haiku 5.5 using high effort for all. Same prompts, same files. Then every result was checked against the real answer.

**Code search** (find every file to change for a new feature):  
\- Files found, out of the 48 the feature really changed: Opus 45, Sonnet 44, Haiku 44  
\- Cost: Opus $6.17, Sonnet $1.38, Haiku $0.44  
\- Same quality, so Haiku wins on price

**Web research** (3 tasks, checked against official rulebooks and Google's help pages):  
\- Wrong facts: Opus 1, Sonnet 3, Haiku 10  
\- On Google Play account rules, out of 12 key facts: Opus got 12 right, Sonnet 9, Haiku 5  
\- Haiku couldn't open a league's official rulebook PDF, so it used last season's rules instead

**Fact-checking** 2 blog posts before publishing:  
\- Real errors found: Opus 8, Sonnet 7, Haiku 4  
\- Correct sentences wrongly marked as errors: Opus 0, Sonnet 6, Haiku 30  
\- If I had followed Haiku, I would have deleted half of a correct post

Cost for all 5 web tasks: Opus $3.93, Sonnet $1.59, Haiku $0.12

What I use now:  
\- Haiku for code search  
\- Opus for research and fact-checking  
\- Sonnet when small mistakes are OK

Has anyone tested Sonnet vs Opus on research? I'd like to know if my results hold up.

document fetched 2026-10-10 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x1e7ts/i_tested_haiku_55_sonnet_55_and_opus_55_as