Personally I don't care about common benchmarks as I don't find actual coding agent performance correlates strongly with them. One reason to use open-weight models is that they don't hide the reasoning, so you very aggressive context management in your harness to use significantly less tokens. Smaller prompts are faster since KV is N^2 plus it can be dramatically cheaper (if you balance your aggressive context management with maintaining the prefix cache as much as possible). Even paying for API prices directly and using agents as much as I want I spend less per month than the $200 I spent on a Claude MAX sub when I had it.

It’s not, you drink too much Kool-Aid, and your employer has too much money. Or the things you do are simply so irrelevant that nobody cares you’re sending stuff to a company of dubious morality.

OpenCode Go provides a lot of usage for these models for the paltry sum of $10/month. Z.ai's coding plan provides a single-digit multiple of Claude Code's usage for a similar price and performance level. Kimi and DeepSeek models are hundreds of billions of parameters (or, in K3's case, >1T). Many of these models have Opus-level benchmarks and, as I pointed out previously, often practically outperform Anthropic models because they're more consistent.

You should run the maths once. Those tokens cost you much more than the hardware would. But yeah, CAPEX vs OPEX something something.