The mechanism sits in where the tokens go. A transformer reads the whole prompt at once, then produces its answer a word at a time, and when it reasons it does that reasoning in language, working through the stages of a problem in strict order [11]. Every stage is billed, because metering is per token and a token is roughly four characters of text, ingested or emitted [13]. Harder problems want longer chains [14], and prompt evaluation grows quadratically, so twice the input costs four times the compute [12]. Pathway calls its alternative a post-transformer architecture [10] and offers the cost figure, not the score, as the result.
Invert the study's ratio and it gets concrete: BDH-CQ meters at about 9% of its comparator, since 1 divided by 11 is 0.091 [1]. The parameter gap is wider than the cost gap. At 150 million parameters [6], against the tens of billions to hundreds of billions typical of frontier systems such as Llama 3 70B and Llama 3.1 405B [7], the model is roughly 467 times smaller than the first and about 2,700 times smaller than the second [2]. Smaller models are cheaper to run and quicker to train [8], which is precisely why the question worth asking is whether the cheapness is architectural or just a side effect of being small.
The score carries footnotes. Almost 30% is a two-attempt figure [3], not a single shot, and plenty of models sit well above it [15]. The comparator's result is characterised only as slightly higher, a "modest accuracy gain" [5], so cost per correct answer cannot be computed from what has been published, and cost per correct answer is what a buyer actually spends. Relative token cost is a ratio inside a metering convention rather than dollars on a serving stack, and as reported there are no latency or throughput figures on named hardware.
ARC-AGI-1 asks a system to induce a rule from a few worked examples of shapes and rotations [4]. That is a clean measurement of one capability, and it is not what drives most production bills, which tend to be long inputs and repeated tool calls. A cheap non-verbal reasoner that has not met those workloads has not been priced against them either.
My read, with its conditions stated: this is an economics claim on a preprint [1], not a capability claim, and the load-bearing part is the untested one, that the same cost profile holds when the architecture is built at larger parameter counts [9]. If it does, Pathway's scientists get the test they want of whether wide adoption would change deployment cost and scale [16]. If the advantage came from smallness, it will thin out as soon as accuracy has to rise.
