Anthropic just dropped Sonnet 4.6. It’s fast and it’s good. But the real question isn’t which model to use — it’s when you’re reaching for the expensive one out of habit.
When you’re deep in a build, Opus feels like the safe default. It handles everything you throw at it. But “handles everything” isn’t the same as “needed for everything.” I was using Opus for tasks that Sonnet would’ve nailed — meeting prep, code review, project planning — because I never set up guardrails to tell me otherwise.
The fix
I added a tier table directly to my CLAUDE.md system prompt:
| Tier | Model | Use For |
|-----------|--------|--------------------------------------------------|
| Haiku | haiku | Lookups, grep, file reads, status checks |
| Sonnet | sonnet | Daily workflows, planning, reviews, writing |
| Opus | opus | Complex architecture, multi-domain synthesis |Then one behavioral instruction: “Actively flag when I’m using the wrong tier for a task.”
That’s it. The model now nudges me when something doesn’t warrant Opus, or when a task genuinely needs the upgrade from Sonnet.
What surprised me
~95% of my daily work runs fine on Sonnet. The tasks that feel like they need Opus don’t. Opus earns its cost for maybe 5% of the work: multi-file architecture decisions, board-of-advisors deliberations, new system design.
The model doesn’t know what it costs. You have to tell it. And once you do, it’s surprisingly good at self-selecting.