同一 prompt 实测：M3.1 Flash 1h06m $0.22，DeepSeek V4.1 Flash 1h28
两个新 Flash 模型跑同一任务，M3.1 Flash 快 22 分钟、便宜 2.4 倍；DeepSeek 烧了 8950 万 token 却只收 $0.54，因为几乎全程命中缓存，作者提醒账单取决于缓存命中率。
DEEPSEEK BURNED 11X MORE TOKENS THAN MINIMAX ON THE SAME PROMPT. IT STILL PAID 54 CENTS. THE CATCH IS THE CACHE. round 1. same prompt, same task, the two new flash models: > deepseek v4.1 flash: 1h 28m, 89.5M tokens, $0.54 > minimax m3.1 flash: 1h 06m, 8.1M tokens, $0.22 minimax won the round. 22 min faster, 2.4x cheaper. but look at that deepseek token count. at the normal input price, 89.5M tokens is $26.85. deepseek charges $0.006 per 1M on a cache hit. 89.5M × $0.006 = $0.54. the exact bill. so the agent spent almost the whole run rereading its own context from cache. the price list isn't your bill. your cache hit rate is: > keep the system prompt and file order stable > don't let the agent rewrite early context mid run > every cache miss bills at full price today deepseek open sourced deepinfer, the engine it says runs v4.1 flash at scale. the whale's kitchen is public now. the cheapest model is the one that remembers what it already read. one prompt, two models, real receipts. a new fight every day.
这条是引用着另一条写的：x.com/hamelhusain/status/2112588257104216294