Luna, and OpenAI more broadly, still wins on overall cost efficiency though. Haiku comes in at 3x+ pricier than Luna ($0.21 vs .07). This prob winds up closer to 4x or 5x once cost adjustments are made for long horizon tasks on Haiku, which are more expensive.
As enterprise clients develop benchmarks for job-relevant tasks, efficiency metrics will become more meaningful, as well as translate with higher fidelity to increased productivity among workers. Anthropic likely still has some catching up to do.