The value is shifting up the stack... make core intelligence free, then monetize the ecosystem built on top of it .. kinda similar to how the internet itself is free, but platforms and apps capture the value.
I think we'll see a huge push towards connectors for work and personal tools, along with much deeper os level integration. That's where the long-term moat is, not the base model itself.
Improving the free offerings is meant to increase visibility and therefore market share. It's not about goodwill, and it never will be. :)
The free stuff is primarily marketing and marketing always has costs.
In terms of compute, I have no way to really look behind the curtain and see what goes on back there. But I know with codex CLI, in terms of weekly quota: I can get a ton of work done with luna and usually get reasonable results. It feels very compute-light in this way.
I'm amazed by the work luna on xhigh can do for the price I pay (just $20, every month). It has the presentation of something that is very efficient to run, while also being something that can actually produce OK results. It's also fairly quick.
It differs from many previous smaller offerings of yore in this way. Like, I mean: I found stuff like the -mini models and 4o to be utterly useless wastes of my time. Luna isn't like that at all; it can get some stuff done.
So far for me, luna is the most impressive part of the 5.6 rollout. Not because it is best, but because it is useful and cheap.
So if luna is decent (it seems that it is), and if it is in fact light (which seems to be true from what I can observe), and it is offered for free, then it may very well be better, faster, and cheaper than the competition is.
And that's good for visibility. Marketing is all about buying eyeballs.