The model stopped being the variable
I kept switching frontier models expecting to find a difference, and the spread turned out to be prose quality and verbosity, not capability. The real gap is judgement, and it is present in every model equally — so the leverage is below the model, in the context and weighting you bring.
I keep switching models expecting to find the difference, and I don’t find it.
Three versions of one frontier model — practically the same result on the task. The newest one reliably adds only verbosity. And between that model and its closest competitor I see no difference in solving the task either; the competitor writes better text and is more pleasant to work with, which is real but is not the same claim. The frontier one needs more attention when it hands you a plan — except I caught the competitor doing exactly the same thing, so that isn’t the differentiator either.
Which means model choice is no longer the interesting decision
If the spread between three versions of one frontier model and a competitor is prose quality and verbosity, then the thing I am choosing between is not capability. It is ergonomics. That is worth something, but it should not get the attention budget it currently gets, and it should certainly not be the axis anyone plans architecture on.
The gap that is actually there is judgement, and it is everywhere equally
They collect facts well. They cannot weigh them.
Ask for research and you get a combine harvester — facts gathered, reasoning arranged around them, genuinely useful. Ask which of it matters and it falls over. It cannot say what is important and what can be dropped, so it fastens onto some incidental feature and chews on it for the whole document. I had one seize on the fact that I have an ingress configured and grind through the entire research on that, when the honest answer was “you can ignore this, move on”.
And the tell that it is not reasoning about importance: correct it once and every conclusion flips to the opposite. Not adjusted — inverted. Nothing was weighed in the first place, so nothing resisted.
So the leverage is below the model
If the models are level on the task and all equally missing the same faculty, then what I add has to be the thing they don’t have: the context, the weighting, the statement of what matters here. That is the harness argument arriving from a different direction, and it is also why the local path is not a downgrade. A ~27B quant at 256k context doesn’t do much yet. But if the frontier’s advantage on my actual work is prose, then the subsidised pricing is buying me less than the bill implies, and that gap closes from below.