I haven't had any issues using OpenAI models (I have a free month, don't worry Sam Altman won't actually be getting my money). But I deliberately use OpenCode for my harness and have an account with OpenRouter so I can switch models in two clicks if one ever started to refuse.

I'm kind of surprised at how most projects seem to tunnel vision onto Claude. So far most of my own REA project was handled just fine with models like DeepSeek, MiMo and Minimax, or by ChatGPT Luna. Occasionally I step it up to Terra for a more complex problem, but it burns through my allocation fast so I don't do it much. There's still two model tiers available above that which are even more capable (and expensive).

And all the benchmarks I've seen suggest that Claude models are much more expensive per task than basically anything else, so it seems silly for the community to be so myopic about it. Especially when Anthropic seem hostile to the whole use case.

Yeah, I'mma disagree. All I see are high level security theats when it comes to using generative AI for reverse engineering. Mainly because vast supermajority of the blooming idiots using it wouldn't even understand a lick of code, let alone what lines leave them vulnerable if they connect to the web.

Cool, as in break a whole industry eventually. Sure. I guess so.