The agent way of doing code generation is light years ahead of the old way of trying to convey error messages back to the LLM manually. You tell ChatGPT5 what you want, it writes code and answers questions, then gives you command lines to try, and watches the output for errors, then fixes them, and has you try again, until it works.
Awesome stuff, affordably priced, even for a retired dude on a very fixed income like me.
You do have to watch out, though... ChatGPT5 apparently knows way more Computer Science than I do, it spit out code in a few minutes that I'm still trying to figure out, days later.[1] (line 245, decompile_lut_to_expr(lut: int) -> str)
It's going to take me another day or two to fully grok the Quine–McCluskey algorithm, and the Möbius transform it uses to take a compiled expression of 4 variables, and regenerate the minimum expression necessary to create it.
[1] https://github.com/mikewarot/Bitgrid_python/blob/main/bitgri...
Had a Gemini Pro sub for 1 or 2 months but it didn't impress me that much.
When I need real horsepower (usually advanced math stuff) I use DeepSeek, which is both free and unparalleled in my opinion.
I did some experiments with OpenRouter but my total usage is still below $10.
More and more I've been experimenting with local thinking models and while it's quite slow (on my non-Apple, CPU-heavy hardware, at least), for some use cases it's an acceptable trade-off and the results have been satisfactory.
I care a lot about data privacy and I'm just not gonna upload clients' proprietary codebases under NDA to some AI company. I've been considering buying or renting proper hardware for local inference but the TCO (particularly with regards to depreciation) is not intuitive.
(mainly keeping the redundancy for availability purposes)
We have a soft cap at $500/mo for everyone right now and most devs don’t get close.
even the free version is enough.
