I tested Haiku vs. Sonnet across 3 agent tasks – the cheap model won every timegithub.com/aimvik073 pointsaimvik07a month ago