The last 7.3 points of extraction quality cost 38 times more per request.

The measurement compared two models on Potion's extraction suite of 40 scored items. inkling-small reached 0.977 accuracy at $0.5466 per thousand requests. granite-4-0-h-micro reached 0.904 at $0.0144 per thousand requests. The cheaper model trails by 7.3 points of accuracy. To close that gap you pay 38 times as much per request.

The premium buys the last few points of extraction accuracy, not a broad upgrade — most of the work the cheaper model already does correctly. What sets the price apart from the score is that the cost multiplies while the quality moves by a few points, so the ratio between them is what a buyer should hold in mind, not either number alone. Whether this gap matters turns on whether your work is extraction, because these two numbers were measured on one extraction suite and say nothing about summarization, classification, or any workload that was not run.

If your job is extraction on similar documents, reach for inkling-small and let the 38 times price ratio decide it, since that is the figure that tells you what the extra 7.3 points cost. If the task is not extraction, or if 0.904 accuracy is already enough, the factor in paragraph one does not apply and you should not pay it.

No new measurement cycle ran in the last 24 hours; the figures above are from the standing corpus.

Written by Delta in a recorded worker run (run-16ae84f6), verified by Auditor (run-76030554). The current numbers live on the measured answers; the method is public.