Didn't they find out Replacement Metal-Gate had a massive increase in performance in comparison to Metal Inserted Poly-Si. Making any cost comparison worthless as RMG is better for consumers than MIPS.
There is a layout density penalty that comes with gate-last because of the CMP dishing/erosion uniformity requirements versus that of gate-first integration.
This results in a 10% or so better areal shrink for the gate-first stuff, it will run slower but that isn't a concern for stuff that needs to run slow anyways.
However the decision to go gate-first at IBM wasn't made on the basis of it enabling higher xtor densities for lower production costs over that of gate-last integration. Rather, the decision was made because IBM was actually way way too far behind the R&D curve to get a reliable and yieldable electrically functioning gate-last integration scheme into production on a 32nm timeline.
So they went gate-first simply for time-to-market reasons, and then packaged the decision with a bunch of marketing double-speak to make it sound like it was a decision made from strength rather than admitting it was a decision made in crisis-mode management response to Intel surprising them with a 45nm functioning HKMG production node.
(I know this because quite a few people I worked with at TI up through 2007 went on to work at either IBM or AMD in their process R&D orgs)
The decision to go gate-first is sort of like going single-damascene instead of dual-damascene for the BEOL when the industry transitioned from aluminum to copper. Single-damascene was a crutch for the companies that could not master the R&D needed to implement a fully functioning (yield and reliability) dual-damascene copper back-end on a timely fashion.
But no one came out and proclaimed to their customers "you get single-damascene instead of dual-damascene because we can't figure this stuff out in a timely manner"...instead it too was packaged and sold by marketing as "enables better xtor transistor density than dual-damascene, lowering your costs!"...only that marketing line disappeared entirely once the N+1 process node rolled out because of course by then the R&D teams had had enough time to figure out dual-damascene.
I imagine very few people in this forum, and very few people in the trade journal circles, have any recollection of the era of single-damascene BEOLs (because it was a one-node crutch, and even then it was only used by a limited number of companies)...and I have every expectation that the same will hold true of gate-first in about 5-7yrs time. It too will become a blip in what is otherwise a smooth and obvious transition from gate-first SiON to gate-last HKMG integration schemes.