Sep 7, 2026
Same 10 in and 50 out per million tokens. One burns 3.3x the tokens per index task and bills 2.4x. One cache line sits 4x apart.