Space Bunny（M3.1 Flash 匿名版）揭晓：作者更关心缓存输入 $0.015/M 的定价
作者认为比 DeepSWE 74% 更关键的是 $0.15/$0.60/$0.015 每百万 token 的输入/输出/缓存价格，并希望看到含重试与验证的「每成功任务成本」，同时期待 10 月 23 日开源权重。
So Space Bunny was MiniMax M3.1 Flash all along! Honestly, the number that excites me most isn't even the 74% on DeepSWE. It's the pricing: $0.15/M input tokens, $0.60/M output tokens, and just $0.015/M cached input tokens. Coding agents aren't one-shot benchmarks. They repeatedly read context, call tools, make mistakes, inspect failures, and try again. Making those iterations faster and cheaper can fundamentally change how much autonomy developers can afford. MiniMax also reports 3x the inference speed at roughly one-third of M3's agent-task cost. That's an exciting combination. Of course, the metric I'd really like to see is cost per successfully completed task, including retries, verification, and wall-clock time. Cheap tokens are nice. Cheap, reliable outcomes are much more interesting. And with open weights arriving on October 23, I'm especially excited to see what the inference community can do with it. Huge congrats to @MiniMax_AI and the team!
这条是引用着另一条写的：x.com/minimax_ai/status/2112150867868541221