I have recently been trying to replace some Rust functions with specifications written in Markdown. Then I can just use an LLM to generate the function according to the specification. I used a local model to generate the code correctly, but I decided to try some online models.
At first I tried GPT-6-Luna. It did not manage to generate correct Rust code. I then tried GPT-6-Sol, and it worked well—it took GPT-6-Sol less than a minute to generate the code for the 2 specs, while my local model took around an hour.
GPT-6-Sol cost me about $0.07 (7 cents) for the 2 code generations I did. Meanwhile, my local model, which ran for about 60 minutes at 100% GPU usage, likely cost me as much as $0.05 in electricity (5 cents). So I did not save much money by using a slower, weaker model at home.
I had GPT-6-Sol generate more optimized code as well, so the task was a success. Getting used to spec-driven development will take some time, but I have had promising results so far—it seems like a big, powerful model can generate code from these specs, while a weaker, local model could be used for maintaining glue code in the rest of the program.