OpenAI launched GPT-6 Sol and GPT-6 Luna today (September 22, 2026). They sit alongside flagship GPT-6 Astra, which shipped earlier this month. There is no GPT-6 Terra at launch; the old cheapest Terra tier is reportedly discontinued.
This is a pricing and tiering story as much as a model story. Below is what matters for teams that actually run volume, and what it doesn’t change.
What shipped
According to coverage including TechCrunch and OpenAI’s community announcement:
- Sol: positioned for complex work and coding. API pricing around $2 / $10 per million input/output tokens.
- Luna: positioned for high-volume tasks: summarize, extract, quick Q&A. API pricing around $0.10 / $0.50 per million input/output tokens.
- ~50% lower API prices versus GPT-5.6 Sol/Luna, with OpenAI citing caching and inference efficiency.
Astra remains the flagship; Sol and Luna fill the “do the work at scale” lanes.
Where you can use them
- ChatGPT Work + Codex for Plus, Pro, Business, Enterprise, and Edu.
- Luna also available on Free and Go via the desktop app.
- ChatGPT web rolling out through the day of launch.
If you’re already on Work or Codex, expect Sol/Luna to show up in the model picker as rollout completes.
Why cheaper high-volume models matter
For operators, Luna-class pricing changes the default on high-throughput jobs: summarizing threads, extracting fields, answering routine questions at volume. Sol-class pricing makes heavier coding and multi-step work less painful to leave on the API meter.
That is real. We use models every day: Grok Bot teammates, Python and web overflow trackers, the usual operator stack. Lower unit cost means you can afford more passes, more batch jobs, more “just try it” experiments without flinching at the invoice.
What cheaper models don’t fix
Models alone don’t clear ops out of email.
Cheaper tokens don’t invent:
- Bounded workflows: clear inputs, clear exits, stop conditions when confidence is low.
- Role-specialized agents: one job per agent, not a single mega-prompt that “does everything.”
- Humans on judgment: exceptions, client-facing decisions, and anything that shouldn’t auto-send.
Without those, you just get cheaper noise: more summaries nobody reads, more extracts nobody trusts, more drafts that still need a human to decide.
We’re not OpenAI partners. We’re operators. The stack that works for us is: right-sized model for the task → agent with a narrow role → workflow with hard bounds → human on the judgment call.
Practical takeaway for teams
- Route by job. Luna-class for volume extract/summarize/Q&A; Sol-class for complex and coding; keep flagship (Astra) where quality ceiling matters more than unit price.
- Meter the workflow, not the model. Track cost per completed outcome (ticket closed, draft approved), not just tokens burned.
- Keep a human gate on anything that leaves the building or changes state in a system of record.
If you’re redesigning how AI sits inside ops (not just which model is cheapest this week), that’s the work we do at vikinglabs.com.
Leave a Reply