Owen Posted on May 10 • Originally published at ofox.ai TL;DR Real LLM bills run 3 to 50 times lower than the headline per-million-token price because most input tokens come from cache. DeepSeek's deepseek-v4-flash cache read is $0.0028 per million versus $0.14 cache miss — a 50x discount. Claude Opus 4.6 cache read is $0.50 per million versus $5.00 input. OpenAI GPT-5.5 cached input is $0.50 versus $5.00 cache miss. If you're paying full price, you're either streaming a moving target into your prompt prefix or your hit rate audit is wrong. The 90% Discount Nobody Calculates Correctly Open any LLM pricing page in May 2026 and the headline number is the cache miss price, not the price you actually pay. If your prompt has a timestamp anywhere near the top, your cache hit rate is zero and you're paying the cache-write tax for nothing. This is the gap behind every confused Slack thread that starts with "we're spending $4 a day, the calculator said $50, what's going on?" Pricing pages quote miss prices.…