I think it depends on where you think we are on the S curve of intelligence growth. (Yes, I think it's an S curve, not an unbounded exponential). If you think we're near the peak than playing catch up (especially if you can play catch up quickly) is very rational.

I know this isn't exactly a scientific test, but I had a local Qwen 3.6 27B model implement a fairly sizable feature today. There were a couple of bugs, mostly around me not giving sufficient specifications, but they were ironed out quickly when I pointed it out. I was able to ask the model to create instructions so next time it doesn't fall into the same pitfalls, and it did a great job. 27B local model! (And it was super fast too).

I ran Fable 5 as a code review and it didn't really have any significant corrections.

I guess my point here is that, for most work the frontier models are probably overkill anyway, and improving on overkill in a way that raises prices significantly is probably not a winning strategy.

The only place I can think of where the super high powered models are "required" is if you want to do a ridiculous token burn like GasTown where you just have it run un-monitored on very long tasks. To me though, that's an experiment, not a real workflow. And the way these labs are like "oh we made this (broken) thing in a week using just agents!" always also follows with "and it cost $100,000+ in tokens!". Like, ok, I get it if you're doing research but that's the salary of an entire person.. that can actually learn and improve.