I tested Haiku vs. Sonnet across 3 agent tasks – the cheap model won every time

(github.com) by aimvik07 | May 21, 2026 | 0 comments on HN