OpenAI adds Sol and Luna under Astra

OpenAI published Introducing GPT-6 Sol and Luna on 22 September 2026. The post adds two GPT-6 family models below GPT-6 Astra: Sol for complex coding and agentic work, and Luna for efficient high-volume tasks.

OpenAI says API prices for Sol and Luna are 50% lower than GPT-5.6 promotional pricing. Docs list Sol at $2 input and $10 output per million tokens and Luna at $0.10 input and $0.50 output. API model ids are gpt-6-sol and gpt-6-luna.

Work, Codex, and API get both ids today

ChatGPT Work and Codex users on Plus, Pro, Business, Enterprise, and Edu plans get Sol and Luna starting 22 September. Free and Go users can use Luna in the desktop app only. OpenAI says the models are not yet available in Chat.

API customers can call gpt-6-sol and gpt-6-luna now. Official model pages document context window (1,050,000), max output (128,000 tokens), reasoning effort options through max, and Responses API tool support.

OpenAI also posted the same launch note on the Developer Community announcements board.

Two mid-tier ids at half GPT-5.6 promo prices

  1. Sol and Luna are two mid-tier GPT-6 ids at 50% below GPT-5.6 promotional list prices. Sol costs $2 input and $10 output per MTok (cached input $0.20). Luna costs $0.10 input and $0.50 output (cached input $0.01). OpenAI frames both as a 50% cut versus GPT-5.6 Sol and Luna promotional rates.
  2. OpenAI places both models under Astra. The blog keeps GPT-6 Astra as the top model for hardest work. OpenAI says Sol and Luna bring Astra's advances in professional work, factuality, coding, computer use, and alignment to faster, more affordable models.
  3. Paid tiers get higher usage headroom in ChatGPT Work and Codex, with a gradual same-day roll-out. Free and Go access is Luna on desktop only.
  4. OpenAI reports eval deltas on AutomationBench, Agents' Last Exam, FrontierCode, DeepSWE, OSWorld, and an internal factuality evaluation. Treat these as OpenAI's own comparisons unless you re-run them.

Anthropic's Claude Haiku 5.5 now matches Luna's $0.10 input and $0.50 output list price for prompts up to 100,000 tokens. See Claude Haiku 5.5 prices like GPT-6 Luna until prompts pass 100,000 tokens.

No Chat date and lab-only competitor rows

The blog's competitor rows (Claude Opus 5, Claude Fable 5 and 5.1) come from OpenAI's research environment or public reports. Effort settings, tool harnesses, and fallback accounting differ by row. OpenAI notes some Fable cost figures omit Opus fallback spend.

Sol and Luna are not in Chat yet. Desktop-only Luna for Free and Go stays a surface limit until Chat availability ships.

Regional processing premiums, Batch, Flex, and Fast multipliers, and prompts over 272K tokens change the invoice. Read the Sol and Luna pricing footnotes before forecasting.

No shipment date is given for Chat availability beyond "gradually throughout the day" for Work and Codex visibility.

PewDiePie says OpenAI banned him twice while he tried to distill a model he calls Sol. See PewDiePie unveils Ajax, a refusal-ablated Qwen 3.5 9B agent for Odysseus, before its weights are out.

Retarget gpt-6-sol and gpt-6-luna traffic

  • API teams on GPT-5.6 Sol and Luna should retarget traffic to gpt-6-sol and gpt-6-luna and re-measure cost per task at the same reasoning.effort. Confirm cache-write billing at 1.25× input.
  • Codex and ChatGPT Work admins should check model pickers after the staggered roll-out. If Sol or Luna is missing, retry later the same day as OpenAI advises.
  • Free and Go users should expect Luna on desktop only. Do not plan Chat workflows on these ids yet.
  • Keep Astra for jobs that already need maximum depth; the blog still points hard work there.

A later article covers OpenAI's GPT-6.1 Sol release on 29 September 2026, which keeps the Sol list prices and cuts cached input to $0.10.