MiniMax M3.1 Flash 发布：AA 智能指数 42，任务成本降至 $0.22
汇总官方规格与 Artificial Analysis 独立数据：智能指数 42（前代 29）、HLE 48%、单任务成本约 $0.22，与 GLM 5.3 Flash 持平、略低于 DeepSeek V4.1 Flash 的 45；并指出其「宁可答错也不承认不知道」的倾向。
MiniMax's new model: nearly 3x faster, 70% cheaper. The Chinese startup released M3.1 Flash yesterday and called the gains generational for agent tasks, coding and cybersecurity. What the independent tests say: • Artificial Analysis scores it 42 on its Intelligence Index, up from 29 for the previous M3 • 48% on Humanity's Last Exam, up from 39% • The average cost to run one of those test tasks fell from about $0.51 to $0.22, cheaper than a cup of chai • That puts it level with GLM 5.3 Flash and just behind DeepSeek V4.1 Flash at 45 The build, as reported: • A 428B parameter mixture of experts model with about 23B active per token • A 1 million token context window • $0.15 per million input tokens and $0.60 per million output • Open weights due on Oct 23 One honest caveat. Artificial Analysis found it has an unusually high tendency to give a wrong answer instead of admitting it doesn't know. That matters most for agents running without a human checking the work. Cheap Chinese models keep getting closer to the frontier on scores. The question is no longer can they compete. It's where you can trust them.