Experimenting with GPT Luna
Two Hacker News comments
After reading a couple of comments on Hacker News, I started reflecting on the latest AI models and how everything has evolved over the last few months.
On the one hand, the latest GPT Luna models and the latest DeepSeek Flash update
are very, very cheap. What catches my attention the most is that using max with these models seems to give them quite a bit of power.
So I looked at a few benchmarks and decided to run a small experiment.
Current benchmarks
In the Artificial Analysis Intelligence Index, GPT Luna Max scores 51 points. It is behind the models at the top of the ranking, but it is already in a fairly competitive range.
The second screenshot is even more interesting because it compares Luna with models that marked a turning point, such as GPT-4.1 when it came out and Claude Opus 4.6, which, at least for me, marked a before and after in using AI for programming.
Seeing them on the same chart is a reminder of how quickly all of this moves.
The experiment
I tested several OpenAI models with exactly the same task and the same starting point and context, to understand whether all the models behaved the same way and whether they could solve it. In the end, I measured time and cost.
It was a relatively simple programming task, but it involved looking up what I was referring to in my prompt, figuring out which files would be involved, making a few changes, and adding tests.
This experiment is more anecdotal than anything else, but the results are still interesting:
| Model | Time | Cost |
|---|---|---|
| Sol Medium | 208 seconds | 47.7 cents |
| Terra High | 67 seconds | 19.1 cents |
| Terra Medium | 45 seconds | 13.9 cents |
| Luna Max | 96 seconds | 2.7 cents |
| Luna Medium | 41 seconds | 1.6 cents |
The price difference is quite significant. Using Luna Max as the baseline, the table looks like this:
| Model | Time | Cost |
|---|---|---|
| Luna Medium | 0.43x | 0.59x |
| Luna Max* | 1.00x | 1.00x |
| Terra Medium | 0.47x | 5.15x |
| Terra High | 0.70x | 7.07x |
| Sol Medium | 2.17x | 17.67x |
Luna Medium is considerably faster and cheaper, although it is probably not capable of handling complex tasks. If you want to save money and time is not a problem, Luna Max is a good model for many things. If cost is not an issue, Terra High seems like a reasonable choice: it is much more expensive, but relatively fast. Finally, Sol is very, very expensive and much slower than the others, although it may certainly be able to do things the others cannot.
In my day-to-day work, Terra has genuinely surprised me. From now on, I will probably use Luna Max more often for personal projects because its price is unbeatable.
I am still not sure to what extent it makes sense, in terms of time and money, to make a first pass on a task with cheaper models, such as Luna Medium, and then use Terra High to review and improve the implementation. But it is certainly worth exploring.
Conclusion
With these cheaper new models and their current capabilities, I’m just hoping in the future we can see models with similar capabilities and prices, but 10x or 100x faster.