GPT-6 Sol and Luna Are Here: My First Impressions
OpenAI released GPT-6 Sol and Luna. The first thing I noticed in the announcement was the price: OpenAI says improvements to caching and inference efficiency let it lower the API prices for Sol and Luna.
Per million tokens, Sol’s input price went from $4 to $2 and its output price from $20 to $10. Luna’s input price went from $0.20 to $0.10, and its output price from $1.20 to $0.50. OpenAI describes this as a 50% reduction compared with GPT-5.6 promotional pricing. Announcement
I also want to thank Chinese models, especially DeepSeek. OpenAI attributes the lower prices to caching and inference efficiency, which is its explanation. But Chinese models started the price competition first, and users can now get cheaper models. I think they deserve some credit for that. There is an affordable GPT-6 Luna now, so I am keeping that thank-you on the books.
Two lines from the announcement worth remembering
“make advanced AI practical for more everyday tasks and applications at scale”
In other words: make advanced AI useful for more everyday tasks and support more applications at scale. —OpenAI announcement
“reducing API prices for Sol and Luna by 50%”
Sol and Luna’s API prices are down 50% compared with GPT-5.6 promotional pricing. —OpenAI announcement
The announcement is about API prices, and the reductions look substantial on paper. I did not notice my own usage getting much cheaper today, though. Maybe I was running too many complex tasks at the same time. The speed difference was easier to notice: GPT-6 Luna Fast feels faster than GPT-5.6 Luna Fast. As for task quality, I do not see much difference from 5.6 Luna.
Terra has always had an awkward place
With GPT-5.6, I used Luna at xhigh for almost everything. I switched to Sol only for a particularly difficult bug or when I needed to break a task down and plan it carefully. I could not think of a situation where I specifically needed Terra.
My rough summary of the models is: Luna quickly writes buggy code, Terra carefully writes buggy code, and Sol writes buggy code with great care. That is not an official description, but it gets the idea across. All three can produce bugs; they just differ in how seriously they approach the task.
At least in the GPT-6 announcement, OpenAI presented Astra, Sol, and Luna, without giving Terra a new place in the lineup. That seems reasonable to me. Luna can handle most tasks, Sol can take the harder ones, and I never knew when I was supposed to choose Terra.
But now that Astra exists, Sol may not keep its comfortable place either. Luna can do everyday tasks, and Astra can take the hardest ones. Will Sol inherit Terra’s old awkward position? I really cannot say yet. I will have to use them for a while and see.