When the grass was greener

I happily installed Claude Code, picked the smartest model, Opus, cranked the effort up to Max, and got to work. Everything was wonderful: Opus did brilliant research, wrote a development plan, then brilliantly implemented it, reviewed it and tested it. Those were the good old days, when the grass was greener and чфthe tokens were cheaper. But little by little I noticed the limits were running out faster and faster, and the kidneys I was selling to pay for it all were, sadly, in limited supply.

So I started thinking about how to optimize this whole affair. The solution was lying right on the surface. Opus writes such good plans, especially if you ask it to write them so that a dumber model could carry them out, that you can just run Sonnet in a parallel window, ask it to do the work by the plan, and then ask Opus to check what came out. Life seemed to be getting better. Then agents arrived, and Opus could now spawn a Sonnet agent to do the work by the plan all by itself. It could even supervise the run and adjust or clarify things on the fly (costs extra, though). It seemed like things were as good as they could possibly get. Why optimize any further?

But tasks and projects don't stand still, and the models devour more and more tokens with every passing day. Even with this setup the weekly limits stopped being enough. I wanted to research more and write ever more grandiose plans (especially once Fable showed up), but Sonnet started eating so much (nobody even remembers Haiku anymore, Haiku is dead) that even a dad working three jobs couldn't keep up with the bill.

Time for a rethink. What now, I thought. How do I keep living, and preferably living well? The expensive Claude tokens should be spent on smart models like Opus and on grand ambitions, not on coding by a ready-made plan. That's when it hit me: why not hand the coding off to some cheap model, one that costs about as much as a cup of coffee and works at Sonnet level. And I happened to be in the right place at the right time: the Chinese developers had not been sitting idle either, and released the wonderful DeepSeek, GLM, Kimi and other models that code quite decently, especially when a beautiful plan has already been written for them.

Pasting plans into a separate GLM window was not appealing; giving up the cozy agent workflow in a single terminal felt like a step backwards. So I had to go deeper. In the end I arrived, as I suspect many of us did, at roughly this scheme: "Opus, write the plan, hand it to a GLM agent to code, then review the result."

In the first four days of this new life GLM chewed through half a billion tokens of coding, all of it inside a ~$40 subscription, without touching a single token of my Claude limits. The limits now go entirely to what Opus does best: research, planning and review.

And this is where I have stopped, for now. The grass is not quite as green anymore, but the tokens are almost cheap again.

Has anyone else settled into a split like this: expensive model for planning/review, cheaper model for execution? What broke for you?