I'll also note that the DeepSeek API seems to be really good at caching and their cached input price is more heavily discounted than most providers at $0.003625 (vs. $0.435 for input cache misses). So, it's hard to spend a lot of money fast with DeepSeek.

I was concerned I would need to do something specific in my dumb agent harness to make caching effective, since I'd read Anthropic's reason for forcing people to use Claude Code in order to use the rolling token usage limits on a subscription was because they could control cache behavior more effectively, but DeepSeek seems to be able to handle caching very effectively for raw API calls.

It's not discounted pricing anymore, it's the regular pricing.