Lab · Field note

Don't fly your kids to school in a G6

My obsession outpaced the Pro plan within two weeks and upgraded to the Max plan, but still kept running out of usage. I like working with the Opus model and hadn't figured out how to put agents to work yet, so everything was happening in one long chat. Not only was I sitting there watching the terminal scroll by, I was setting my usage on fire with simple things that could have been handed off to more basic models.

Why

My 5-hour usage windows were getting eaten up in half that time. Getting 2.5 hours of flow when 5 is available was frustrating. Once I started deploying sub-agents for the mechanical stuff, I could not only run multiple parallel sessions, but even with two active, I was hitting the cap less.

I'm also building discipline for when the all-you-can-eat token buffet ends. Don't use a model if a script will work, and don't use Opus when Sonnet will work for a fraction of the cost.

How it works now

I treat Opus like the manager. It's responsible for taking my direction, working with me to figure out what to do, and orchestrating it. Most of what I do involves the models writing code, and Sonnet is a senior engineer that takes instructions well. Haiku is the intern, and gets to do fun things like 100 parallel Google searches. Opus runs the session in auto mode and only comes back to me when it needs my input.

What the data told me

Claude's built in /usage command tells you how much you would have spent based on model selection, but I wanted to see this result over the long term. Unfortunately with my current telemetry I couldn't get it. My logs come back 95% Opus, and the sub-agent burn barely shows up — the tooling doesn't capture it.

So I can't put a number on it yet, but I'm working on that. Even without it, I have a feel for how each model actually behaves, and have built a healthy reflex for asking "could this run on a cheaper model?"

Like almost everything in business, if I were starting over I'd start tracking the data earlier and would stratify my model use. Opus for anything that requires creativity or "figuring it out." Anything my children can do, I give to Haiku. Everything in between goes to Sonnet.

The experiments

  1. Jul 9, 2026

    Tried to measure my own model mix

    Hypothesis: I've pushed enough work to Sonnet/Haiku that it'd show up in the numbers. What happened: my transcripts don't record sub-agent usage, so the mix reads 95% Opus — mostly the orchestrator re-reading context. Result: can't see the thing I most want to see. Need to build the reporting.

  2. Mar 25, 2026

    Night shift (walk-away autonomous runs)

    The idea: set a job going, walk away, come back to finished work — the "night shift," a term I got from Jamon Holmgren. Honestly: a bust so far — I haven't been able to step away as much as I'd like. Could be me, could be the harness. Result: unresolved; still figuring out whether it's a workflow problem or a tooling problem.