The original headline is: "GPT-5.6 Sol vs Terra vs Luna: which tier should you actually use?"

Short summary

GPT-5.6 ships as three tiers — Sol ($5/$30), Terra ($2.50/$15), and Luna ($1/$6) per million tokens. The author ran real traffic across all three and found Terra sufficient for everyday coding work at ~97% of Sol's benchmark performance. The key insight: prompt prefix caching (10% of input price for cache reads) can make Terra-with-cache cheaper than Luna-uncached for prefix-heavy RAG workloads, reframing tier selection as a prefix-ratio modeling problem rather than a simple cost-per-token decision.

• •Terra hits ~97% of Sol's benchmark performance at half the price — sufficient for most daily dev work
• •Prompt caching bills cache reads at 10% of input price, rewriting cost math for prefix-heavy workloads
• •Luna is best for high-volume, low-complexity tasks (classification, tagging) where per-token cost dominates

Generated with AI, which can make mistakes.