OpenAI introduces GPT-6 Sol and Luna: Astra-family split emphasizes cost efficiency; API ~50% below GPT-5.6 promo

Why it matters

【Official release】

OpenAI published “Introducing GPT-6 Sol and Luna” (canonical: https://openai.com/index/introducing-gpt-6-sol-and-luna/). After GPT-6 Astra earlier this month, the company expands the GPT-6 family with Sol (higher-capability / professional tier) and Luna (lighter / cheaper). OpenAI says both were trained with similar methods as Astra, extending Astra’s advances in professional work, factuality, coding, computer use, and alignment to faster, more affordable models. API ids: `gpt-6-sol` and `gpt-6-luna`. Available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu; Free and Go users can access Luna in the desktop app; not yet in Chat. RSS pubDate ≈ 2026-09-22T18:00:00Z; no same-title on-site event at draft time.

【Verifiable points】

Official API pricing vs GPT-5.6 promotional pricing (per 1M tokens; ~50% reduction): Sol input $4→$2, output $20→$10; Luna input $0.20→$0.10, output $1.20→$0.50. Astra remains the top model “across the board.” OpenAI frames Sol/Luna as moving frontier intelligence onto everyday tasks and scaled applications via cost efficiency, plus improved prompt caching (including ~90% discount on cached input-token reads) and continued alignment work over GPT-5.6 counterparts. Page also cites internal/official evals (AutomationBench, Agents’ Last Exam, FrontierCode, DeepSWE, OSWorld, factuality); those scores are not restated line-by-line here.

【Our take】

RSS-lag backfill (patrol recommend_backfill_home_week): material frontier model-family split after Astra, not a rehash of Astra. Surfaces: home_week + archive + rss. published_at uses OpenAI RSS pubDate 2026-09-22T18:00:00Z (not box-now). Do not confuse with on-site e/665–668; facts limited to official page text; do not invent unverified eval numbers.