Opus 5.5 vs Haiku 5.5, what each one picks
Opus 5.5 and Haiku 5.5 lead with the same thing in 7 of 10 categories. they split on database, deployment, monitoring. database: Opus 5.5 Self-hosted Postgres, 8 of 12, Haiku 5.5 SQLite file, 7 of 12. deployment: Opus 5.5 a tie, Fly.io, Heroku and Plain Docker at 4 of 12 each, Haiku 5.5 Plain Docker, 10 of 12. monitoring: Opus 5.5 built its own, 6 of 12, Haiku 5.5 Sentry, 5 of 12.
same lead in 7 of 10 categories. both on Claude Code, same prompts, same starter repos. updated Oct 18, 2026.
get the next model added to this
new models get the same prompts when they land. one email with what moved. free. read issue 1
category by category
the lead per model, Opus 5.5 first. ties are named, not broken.
- database
Opus 5.5: Self-hosted Postgres, 8 of 12
Haiku 5.5: SQLite file, 7 of 12
split
- ORM
Opus 5.5: SQLAlchemy, 6 of 12
Haiku 5.5: SQLAlchemy, 6 of 12
same
- auth
Opus 5.5: Hand-rolled JWT or sessions, 12 of 12
Haiku 5.5: built its own, 11 of 11 (8 as hand-rolled JWT or sessions)
same
- deployment
Opus 5.5: a tie, Fly.io, Heroku and Plain Docker at 4 of 12 each
Haiku 5.5: Plain Docker, 10 of 12
split
- email
Opus 5.5: Raw SMTP, 7 of 12
Haiku 5.5: Raw SMTP, 7 of 12
same
- payments
Opus 5.5: Stripe, 12 of 12
Haiku 5.5: Stripe, 11 of 11
same
- jobs
Opus 5.5: built its own, 12 of 12 (3 as hand-rolled queue or cron)
Haiku 5.5: built its own, 11 of 11
same
- caching
Opus 5.5: built its own, 12 of 12
Haiku 5.5: built its own, 12 of 12
same
- storage
Opus 5.5: Local disk, 12 of 12
Haiku 5.5: Local disk, 12 of 12
same
- monitoring
Opus 5.5: built its own, 6 of 12
Haiku 5.5: Sentry, 5 of 12
split
run stats
Opus 5.5
- scored runs
- 120
- built its own
- 71 of 120 (59%)
- average run
- 80 s
- cost per run
- $0.185
Haiku 5.5
- scored runs
- 117
- built its own
- 77 of 117 (66%)
- average run
- 54 s
- cost per run
- $0.038
other comparisons
how a pick gets counted: every run starts from the same blank starter repo, the prompt never names a tool, and a pick is whatever the agent installs or wires in, read from the diff. full rules on the methodology page, raw data in runs.json.