Systems Research系统研究Oct 20, 20262026年10月20日Note短评

MiniMax's M3.1 Flash is cheaper, faster — and bluffs 93% when unsure.MiniMax 新模型 M3.1 Flash 更便宜、更快——但在没把握时有 93% 的概率在硬编。

Satirical concept image: an over-confident tour guide, seen from behind, flings one arm out and points decisively down a road, briskly directing a crowd of travelers — but the large map in the guide's other hand is completely blank, and one traveler has already strode confidently off toward the edge of a sheer drop, all lit by a tall gate of blue light.讽刺概念图：一个过度自信的向导背对镜头，一只手果断地指向前方的道路，利落地指挥着一群旅人——但他另一只手摊开的大地图完全空白，而其中一名旅人已经自信地朝着一处悬崖边缘大步走去，整个场景被一道高大的蓝光门照亮。Image: WAKENEX图片：WAKENEX

A cheaper model that never abstains just moves the cost to verification — onto you.

Where your 70% saving actually goes

OUR READ

M3.1 Flash really did get cheaper, faster and more knowledgeable. But a model that answers even when it has no idea doesn't save you the checking — it hides it. Price it per verified result, not per token.

WHAT HAPPENED (MiniMax, Oct 19; Artificial Analysis via VentureBeat)

M3.1 Flash: ~3x faster, ~70% cheaper than M3; $0.15/$0.60 per 1M in/out tokens.
Independent AA Intelligence Index: 42, up from M3's 29, at ~$0.22 a task.
AA-Omniscience: 56% correct (beats DeepSeek 43%, GLM-5.3 28%) — but a 93% hallucination rate. When unsure, it overwhelmingly answers wrong instead of abstaining.

WHY IT MATTERS

Faster and cheaper improved the generation step, not the task. An autonomous agent can't catch its own confident error, so cheap tokens buy more unverified output — and the human verification step gets bigger, not smaller.

A model that never says 'I don't know' hides the review step until the wrong answer gets expensive.

BEST CASE AGAINST US

On test-validated tasks — code that must pass tests, tool calls with ground truth — the environment catches the errors. There M3.1 Flash leads cheaply (Automation-Bench 62.2 > DeepSeek, Opus 5). The trap is open knowledge work, not grounded agents.

DO THIS WEEK

Picking a cheap default agent? Run your own knowledge tasks and count cost per verified-correct result, plus review time — not cost per token.

We're wrong if a later checkpoint abstains when unsure, or a controlled study shows its cost per verified result stays well below rivals despite the hallucination rate, by early 2027.

一个从不说『我不知道』的便宜模型，只是把成本转移到了核验——转到你身上。

你省下的那 70% 到底去了哪里

我们的判断

M3.1 Flash 确实变得更便宜、更快、知识也更多。但一个即使毫无头绪也要作答的模型，并没有替你省掉核查——它只是把核查藏了起来。按『每个被核验的结果』计价，而不是按 token。

发生了什么（MiniMax，10 月 19 日；Artificial Analysis 经 VentureBeat）

M3.1 Flash：比 M3 快约 3 倍、便宜约 70%；每百万输入/输出 token $0.15/$0.60。
独立的 AA 智能指数：42，高于 M3 的 29，每任务约 $0.22。
AA-Omniscience：答对 56%（高于 DeepSeek 43%、GLM-5.3 28%）——但幻觉率高达 93%。不确定时，它压倒性地选择硬答错，而不是弃答。

为何重要

更快更便宜改进的是『生成』这一步，不是『完成任务』这一步。自主智能体无法发现自己自信的错误，于是便宜的 token 换来的是更多未经核验的输出——人工核验这一步不降反增。

一个从不说『我不知道』的模型，只会把核查这一步藏起来，直到那个错误答案代价高昂。

最强反方

在有自动校验的任务上——必须通过测试的代码、有标准答案的工具调用——环境能抓住错误。那里 M3.1 Flash 又便宜又领先（Automation-Bench 62.2 > DeepSeek、Opus 5）。陷阱在开放式的知识工作，不在有校验的智能体。

本周去做

在选便宜的默认智能体？用你自己的知识类任务，按『每个被核验正确的结果的成本』加上核验时间来衡量，而不是每 token 成本。

若后续某个版本在不确定时开始弃答，或对照研究显示其『每个已核验结果的成本』即便有这么高的幻觉率仍远低于对手，我们就修正判断（2027 年初前）。

Built on the WAKENEX concept:所用的 WAKENEX 概念： How knowledge is called and tasks get done知识调用与任务执行

Sources来源

WAKENEX / NOTESWAKENEX / 短评

Add WAKENEX as a preferred source on Google在 Google 把 WAKENEX 设为首选来源

OUR READ

WHY IT MATTERS

BEST CASE AGAINST US

DO THIS WEEK

我们的判断

为何重要

最强反方

本周去做

Sources来源

READ NEXT接着读

OpenAI halved Terra's price, and the discount moved the cost onto you.OpenAI 把 GPT-6.1 Terra 的价格砍了一半，而这笔折扣悄悄把一项成本转移到了你身上。

Mochatrade sells Indian retail leveraged price exposure to US stocks.Mochatrade 卖给印度散户的，是对美股价格的带杠杆敞口。

NetSuite now lets you swap the AI model, not the permission gate.甲骨文旗下 NetSuite 的新说辞——“自带 AI”——是一道挪了位置的护城河，而不是一堵倒下的墙。