Yesterday, OpenAI introduced GPT-6 Sol and GPT-6 Luna. After Astra, I was expecting most of the conversation to be about capability again. Instead, the part that stayed with me was efficiency.
These are the new Sol and Luna models in the GPT-6 family. OpenAI says their API prices are lower than the promotional prices of their GPT-5.6 counterparts, with improvements in caching and inference helping reduce the cost. The announcement has the actual comparisons.
Why this matters to me
When I’m learning, I rarely get something right on the first try. I change the question, fix the code, find another mistake, and run it again. A tool that’s useful only once is different from one I can afford to use while figuring things out.
So I’m interested in the cost of reaching a result I can trust. If a cheaper model needs ten corrections, the saving might disappear. If a faster one can handle an ordinary task well, I might not need the biggest model for that task at all.
I’d like to compare them on something small: organising experiment notes, writing a plotting script, or helping me understand an error. I could check whether the output works, how much I had to change, how long it took, and what the whole attempt cost. I haven’t run that comparison yet.
A connection to building robots
This reminds me of a question that comes up with physical systems: how much can I do with the resources I have? A robot has limits on power, space, weight, and computing. More capability is useful, but it has to fit those limits.
That doesn’t mean a cloud language model belongs inside a fast motor-control loop. The timing and reliability requirements are very different. I’m thinking more about supporting tasks, such as reviewing logs or helping me make sense of a test after it finishes.
I’m excited to see models become more practical to use repeatedly. For a student with plenty to learn, being able to try, check, and try again matters quite a lot.