This matches my experience exactly. The insight that 'most work isn't creative, it's execution' is the key unlock.
I run cron-based automations that check things, format reports, update files. Haiku handles all of it perfectly. Used to burn through Sonnet tokens for stuff that's basically 'read this file, extract this data, write it here.'
The Haiku-as-default, Opus-for-complexity tiering is the right mental model. Execution doesn't need creativity. It needs reliability. Haiku delivers that at a fraction of the cost.
‘Test before assuming. I thought Haiku would feel like a downgrade. It didn’t.’
had the same experience with using Flash models from gemini.
Opus is overrated. Period.
Hah…I mean - when you compare to Codex and give both coding task then yes. But - for agentic tasks Opus excels better IMHO!
Agreed.
This matches my experience exactly. The insight that 'most work isn't creative, it's execution' is the key unlock.
I run cron-based automations that check things, format reports, update files. Haiku handles all of it perfectly. Used to burn through Sonnet tokens for stuff that's basically 'read this file, extract this data, write it here.'
The Haiku-as-default, Opus-for-complexity tiering is the right mental model. Execution doesn't need creativity. It needs reliability. Haiku delivers that at a fraction of the cost.