What's this about?
OpenAI brought Astra-generation training improvements to two cheaper tiers. Sol targets complex coding and professional work; Luna targets inexpensive high-volume work.
What changed economically?
Sol is $2/M input and $10/M output tokens; Luna is $0.10/M and $0.50/M. Better caching can discount repeated input by 90%, so agent architecture now matters almost as much as list price.
What's the community saying?
HN 1,251 points, 629-plus comments; official X post 45,000-plus likes at fetch time.
- The 50% cuts are seen as a direct frontier-model price war.
- Users want independent cost-per-completed-task comparisons across harnesses.
- Luna's very low price may matter more to product builders than a small benchmark lead.