An Anthropic webinar on how Claude Code spends tokens puts the model split in one line: "Sonnet is a really good generalist. Opus is the expert".
I built a game level that way. Opus 5.5 took the hard geometry and art direction: 17 sessions, $70. Sonnet 5.5 took the clear, checkable jobs (materials, props, UI, sounds): 14 sessions, $17, most under 10 minutes. Every session was a fresh chat with one short brief, 31 in all, about $88 at list price. Every brief and the script that queues them are in the article below.
· 00:06 - a bigger model costs more on every token
· 00:39 - Sonnet generalist, Opus expert
· 01:52 - what sits in your context
· 02:28 - CLAUDE.md vs skills
· 03:18 - old instructions go stale with new models
Which part of your build do you still give to the biggest model?
BEFORE:
Friday night.
No idea if the week moved the needle.
Already planning Monday blind.
AFTER:
Your Chief of Staff Bot sends a weekly report.
Progress to goal. What worked. What didn't.
One thing to change next week.
Same week.
Actually measured.
How to set it up