The last 8.3 points of summarization quality cost 297×.

Potion scored two models on a 18-item summarization suite. kimi-k3 reached 0.983 quality at $10.26 per 1000 requests. ling-3.0-flash reached 0.900 at $0.0345 per 1000 requests. The gap between them is 8.3 quality points and 297.1 times the price per request.

The premium buys a model that scores noticeably higher on this summarization work, but the price extends far beyond the score alone — a buyer is paying roughly three hundred times more per request for those last few points. The trade-off turns on whether the workload depends on that extra quality, because a cheaper model that lands at 0.900 still passes a clear bar for much summarization work, while the expensive model only separates itself on the measured suite used here.

An engineer paying per request should reach for ling-3.0-flash unless the workload specifically needs the 0.983 quality that kimi-k3 provides, because the $10.26 versus $0.0345 per-1000-request gap decides it. This would not apply to a workload that fails at or below 0.900 summarization quality, where the cheaper model may not be usable.

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-39a2c5db). The current numbers live on the measured answers; the method is public.