Cheaper, and built on the same brains
Both models come from the same generation as GPT-6 Astra, the model OpenAI put out earlier this month and calls its smartest and most aligned yet. Sol and Luna were trained with similar methods but tuned for cost rather than peak capability, and OpenAI says they carry Astra's gains in professional work, coding, computer use and factuality down into cheaper tiers. The price drop is the real headline. Both are 50% cheaper than GPT-5.6's promotional rates. Sol now runs about $2 per million input tokens and $10 per million output, down from $4 and $20, while Luna is far cheaper still at $0.10 and $0.50 per million. Astra stays the one OpenAI points to when you can't compromise on performance.
The numbers OpenAI is leaning on
On a test of real business workflows, OpenAI says Sol beats Anthropic's Claude Opus 5 while costing just 9% as much per task.
That's the comparison the company wants people to notice, and it kept going. On a harder benchmark for long, multi-step professional work, Sol scored 56.4% at its highest effort, which OpenAI says edges out Opus 5's best result at 60% lower cost per task. On coding it says Sol now matches Anthropic's Claude Fable 5.1 on a test of whether AI code is actually ready to merge, and lands within about a point of Fable 5's best on another while costing roughly 80% less. It also claims Sol makes about half as many factual errors as the model before it, based on its own review of conversations where users had flagged mistakes.
Who gets them, and the catch
Sol and Luna are rolling out today inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users can reach Luna through the desktop app. Neither is in the standard chat yet, and developers can call them through the API as gpt-6-sol and gpt-6-luna. The catch worth remembering is that every one of these wins is OpenAI's own number, run on OpenAI's own tests, so the real proof is in daily use. Both ship with the company's latest alignment work, including the guardrails around self-improving AI it made a point of weeks ago, which is the part the price war keeps pushing to the background.