AI API 降价潮：DeepSeek V4.1 Flash 托管商增至 29 家
汇总本月 API 降价：OpenAI GPT-5.6 Luna 降 80%、Opus 5.5 定价 4/20 美元，DeepSeek V4.1 Flash 在 OpenRouter 的托管商从 13 家增至 29 家；并指出编程 agent 的 token 花费 96–99% 是缓存读取。
AI API prices went down again this month, and the "they will have to raise prices someday" theory keeps losing. Max Woolf collected the moves: 1. OpenAI cut GPT-5.6 Luna by 80%, to $0.20 in and $1.20 out per million tokens. 2. Anthropic priced Opus 5.5 at $4 / $20 and says it needs less compute to serve. 3. DeepSeek V4.1 Flash is now served by 29 providers on OpenRouter, up from 13. 4. Kimi K3.1 Pro is close to Opus level at about half the price, with open weights due Nov 4. why they fall: when the same open model runs on 29 hosts and switching is one line of config, the price goes toward the cost of serving it. competition does that, not charity. the number most people miss: for coding agents, cache reads are 96-99% of the token spend. the headline input price matters less than the cache price and how well your agent reuses context. what i do with this: 1. send the easy steps to the cheap model, keep the big one for planning and review 2. keep the context stable so the cache actually hits 3. re-check the model choice every month, today's cheap tier beats last year's flagship