> I've spent several hundred thousand dollars in API costs in the last few years to power my app

Which is kinda sad, if you've used Anthropic products, because you could get comparable performance from cheaper models for vast majority of tasks.

> No, they don't. That's the whole point.

Bad for them; do not use Anthropic then. Besides "the whole point" of conversation you have interrupted is that "low temperature and tight sampling produces stiff boring cliche prose" - which is truism, as those settings control the entropy of the output. And it is utterly irrelevant frankly if one has access to the Claude sampler through API or not; as Anthropic has apparently locked the sampler at very conservative temperature (probably as low as 0.2), there is no way to squeeze god prose out of it, as the logits have been severely messed up wrt to the actual word distribution in standard English.