The graceful-degradation point is the strongest critique here, routing to a cheaper model instead of stopping entirely when a quota hits is a real product decision other providers have made, so it's a fair thing to want changed, not just a complaint about limits existing.

The marginal-$20 framing is the sharper economic insight, and it generalizes past AI subscriptions, "what does this add to what I already have" gives a very different answer than "is this good" once your stack already has overlapping tools.

One pushback: the benchmark table cuts both ways more than the piece sits with, Claude's lead on DeepSWE/FrontierSWE is nontrivial in the same table where Kimi's lead is smaller. A workload concentrated in Claude's strong areas reaches a different marginal conclusion than a general-purpose one, which the piece acknowledges but somewhat undersells.