OpenAI Ships Three Models at Once: GPT-5.6 Sol, Terra, and Luna Are Live Today

TL;DR: OpenAI launched GPT-5.6 publicly today in three stable tiers — Sol ($5/$30/Mtok, flagship), Terra ($2.50/$15), Luna ($1/$6) — after a 30-day government security review, with explicit prompt-cache breakpoints that matter for RAG and agent apps.

After OpenAI submitted GPT-5.6 for government review on June 9, the Department of Commerce cleared it for broad release today. The three-tier naming (Sol, Terra, Luna) is designed to be stable: each tier advances independently on its own cadence rather than getting replaced by a version number, so "Terra" will always mean "the midrange option."

Terra is the key number for production teams: it targets GPT-5.5 performance at half the cost ($2.50/$15 vs. $5/$30). Benchmark it before the end of July if you're currently paying for GPT-5.5.

Developer-side: GPT-5.6 introduces explicit cache breakpoints — set manually, not just auto-detected — with a 30-minute minimum guaranteed cache window. Cache writes cost 1.25x uncached input; reads are 90% off. For agent apps with large shared prefixes (system prompt + retrieval context), this makes caching far more predictable than automatic detection.

Why it matters: Terra's 2x cost reduction at GPT-5.5 performance is a practical win for any production workload. Stable tier naming means you can pin "Terra" without chasing model-ID changes.

Sources: Previewing GPT-5.6 Sol — OpenAI · GPT-5.6 pricing explained — eesel