I do linux kernel, compiler, and operating system dev with Qwen3.5 122b running locally on a Strix Halo 128G ad 35t/s. Pretty much the most complex software problems one can work on.

I think a lot of people just want to put in a credit card and press an easy button.

Yeah the easy button, if translated to a more capable model that requires less hand holding, manual correction and consistently produces better quality code, is of course the point. You wouldn't want to go from Qwen3.5 122b back to GPT 3.5 for coding assistance.

People can definitely be productive with less powerful models. Supermaven or Cursor's tab autocomplete models from a year ago were already a huge boost over the pre-AI days. They just don't have the same capabilities as the leading models.

Curious if you've tried Gpt 5.4 High through Codex to compare for your use case?