you could do something like rent GPU time yourself, and use it to run a higher-quality local model (e.g. one of the Chinese "close to frontier" ones). Not guaranteed to preserve privacy of course, but it at least avoids directly sending the data to OpenAI/Anthropic.