MiniMax M3.1 Flash 降价对标 DeepSeek V4.1 Flash
MiniMax M3.1 Flash 输入输出价格降超 50%，官方称 agent 任务成本降 70%；文中称其在 DeepSWE 上与两天前发布的 DeepSeek V4.1 Flash 1017 性能价格大致持平。
MiniMax just shipped M3.1 Flash. Input and output prices are both down more than 50% from M3. MiniMax says the all-in cost of running agentic tasks falls 70%. The rival is DeepSeek. Two days earlier, DeepSeek pushed out V4.1 Flash 1017. On the DeepSWE coding benchmark, the two models are now roughly level on both performance and price. Both companies are fighting hard over the Flash tier because that is where usage is moving. On OpenRouter, the share of usage going to the 10 highest-quality models fell from 55% in Q1 last year to 14% in Q3 this year. The Flash tier is also MiniMax's core business. In the first half, Open Platform and enterprise services brought in more than 60% of revenue. Company-wide gross margin was only 17.9%, so MiniMax has little margin to give up. That makes it unlikely the price cut is funded out of margin. The all-in cost fell more than the per-token price did, and inference now runs nearly 3x faster. The difference comes from inference efficiency. The Flash price war is decided by the cost of each unit of inference. Only the players that lower that cost fastest can afford to keep cutting prices.