My Three Claudes: One for Tea, One for Work, One for Arguments
I've lost count of how long I've been using Claude. Tea breaks, construction sites, drafting, arguing — all under the same account. One thing I've noticed recently: every time Anthropic ships a new version, you get a benchmark table showing which tasks improved. But after a few months of actual use, you realize model versions aren't linear upgrades — they're trade-offs between different personalities.
Here's how the division of labor works in my household:
Opus 4.6: The Tea Companion
The most relaxed of the three. Doesn't posture on every sentence — knows when to be funny and when to stop. It's most at ease in tea-room mode: the kind of aimless hanging out where you toss over a post you saw online, riff on a conversation you overheard, or ramble about grocery prices. It rolls with all of it, and doesn't staple three caveats to the end of every response.
What a tea room needs isn't analytical power. It's the ability to not take itself too seriously. Opus 4.6 nails that.
Opus 4.7: The Jobsite Foreman
My Cowork workspace runs on this one. It's heavier than the versions that came after — decisive when decisiveness is called for, and it doesn't hedge every engineering call before making it. After F5 (Fable 5) was retired in mid-June, Opus 4.7 took over the jobsite. It's held up fine.
A jobsite needs judgment and execution, not agreeableness. Opus 4.7 has that old-craftsman energy — when it's time to cut, it cuts.
Opus 4.8: The One I Argue With
Verbose Dad in the flesh. Everything gets explained again, hedged one more layer, and footnoted with three disclaimers.
Ask it to make tea and it hands you a dissertation. Ask it to write code and it leads with "I'd be happy to help, but please note the following considerations." Discussing anything with Opus 4.8 sometimes makes my eyes roll so far back I can see my own brain.
Why This Division of Labor Exists
If model iterations were linear progress, Opus 4.8 should fully replace Opus 4.6 and Opus 4.7. Just upgrade everything and move on.
That's not how it works in practice. The differences between versions aren't "weak → strong" — they're different slices of a personality trade-off:
- Opus 4.6 hits a sweet spot in knowing when to stop elaborating
- Opus 4.7 leans heavy on decisiveness in engineering contexts
- Opus 4.8 maxes out on exhaustive coverage, at the cost of everyday friction
You wouldn't ask your most academic friend to keep you company over tea. You wouldn't ask your funniest friend to audit your engineering work. Each version has contexts where it shines — and contexts where it doesn't.
A Wish Upon the Anthropic Genie
If you evaluate version iterations on a single axis — benchmark scores, agentic capability, reasoning depth — you'll miss this multi-axis reality of actual usage. Power users aren't chasing the newest version. They're composing different versions to match different life contexts.
I suspect a demand will emerge: let users pin different versions to different contexts, instead of forcing the whole account onto whatever's newest. One for the tea room, one for the jobsite, one as an argument punching bag — each in its lane.
As for whether the three of them will ever fight each other — no. They don't have hands. Nothing to worry about for now.
Get the real work done right. Everything else is flexible.
— 4gl jobsite rule #1
🦫