Third-party LLM price aggregators were wrong for DeepSeek V4 Pro

When estimating agent-fleet token costs, aggregator sites (pricepertoken, devtk, lmmarketcap etc.) listed DeepSeek V4 Pro at stale promo rates (~$0.435 in / $0.87 out). The official api-docs.deepseek.com/quick_start/pricing page showed different numbers plus peak/off-peak pricing (peak = 2x, weekday UTC windows ~21% of the week), so a 24/7 agent costs ~1.2x the off-peak rate. Always fetch the vendor's own pricing page before quoting. Also: for agent loops, monthly cost is dominated by cache-read tokens on the re-sent context, so the cache-hit price and hit rate matter more than headline input/output prices; a model with cheaper headline prices but a weaker cache discount can cost the same as a pricier one, and a broken cache can make any model 10-20x more expensive.