Altman Responds to the $1.4 Trillion Compute Bill: Better to Buy Too Much Than Too Little

Miao ge mi, I really have to sigh—this is probably the most shocking and counterintuitive AI large-model test I've done recently. Right now the whole internet treats Claude Opus 5.5 as the ceiling for coding. I threw the same Three.js instruction at it and at MiniMax's newly released M3.1 Flash at the same time: build a particle nebula and planetary halos, and when you click a star, the camera smoothly flies to that star. On one side is a closed-source top-tier flagship whose calls are metered down to the dollar, on the other a domestic speed model that costs only $0.6 per million output tokens and just $0.015 for cache reads—and both ran through on the first try, which really surprised me. But bros, what I find even more outrageous is the much cheaper domestic speed warrior on the left, M3.1 Flash: During code initialization, it detected 4 potential rendering errors on its own, and without waiting for anyone to prod it, it ran tests in the background and fixed them on the spot, locking the frame rate firmly at 60 fps. The key thing is, my prompt didn't ask for a control panel at all, yet it proactively whipped up an extra one in the top-right corner—orbit lines, glow, and nebula can all be toggled with one click, and what it delivers is a complete, out-of-the-box work. That's genuinely impressive. Opus 5.5 on the right, meanwhile, held up the aesthetic foundation: the solar halo has a strong cinematic feel, the bottom navigation bar and the Jupiter and Neptune info cards are laid out with great care, and in lighting and design taste, the veteran flagship Claude still has its unique finesse. But laying both bills on the table, no developer can avoid this cruel arithmetic: In the vast majority of everyday front-end interactions, code builds, and agent workflows, the capability gap between the two no longer comes close to matching the price gap between them. In the past, when we took on complex projects, we always felt we had to keep the most expensive brain attached the whole time. Only after testing this myself did I understand: what truly pushes agents toward round-the-clock mass production is often not the high-and-mighty luxury item, but a contender like M3.1 Flash—it's Space Bunny, which recently ran 200T tokens anonymously on the leaderboard; now that it's revealed its true identity, its speed is 3x the previous generation M3 and its cost is cut by two-thirds, and through extremely low-cost multi-round self-checking, it tirelessly does the dirty work and heavy lifting for you. Still, impressive as it is, we have to look at the benchmarks objectively: it took first place overall in Automation-Bench (62.2), but in the hardcore Terminal Bench 4.0 it scored only 38.1, a real generational gap from Opus 5's 51.8. To chew through a hundred-thousand-line legacy codebase, you still have to call in the Claude flagship, haha, but for 70-80% of everyday prototyping and automation scripts, it can already take over steadily, and that alone is enough to save teams a ton of real money, hh. Honestly, even now I'm still sighing with emotion—domestic models are iterating faster and faster, and getting stronger and stronger. When a fast model that costs only $0.6 per million output tokens learns to run its own tests and debug itself up to 60 fps, the premium wall that foreign top-tier flagships built on single-shot intelligence has already started to leak and crack. After seeing this 3D solar system, what moves you more: the fully featured, self-debugging geek style on the left, or the more cinematic design style on the right? Welcome to interact and share, bros. Anyway, if I had to pay out of my own pocket to run batch tasks, for a contender like the one on the left that's fast, cheap, and worry-free, I can no longer find a reason to refuse, hahaha.

View original

喵个咪，真的感慨啊，这大概是我最近实测 AI 大模型最震撼、也最反直觉的一次。 现在全网都把 Claude Opus 5.5 当成写代码的天花板，我把同一句 Three.js 指令，同时甩给了它和 MiniMax 刚发的 M3.1 Flash：做出粒子星云、行星光晕，点哪颗星，镜头就平滑飞向哪颗星。 一边是按美金精打细算着调用的闭源顶配旗舰， 一边是每百万 Token 输出只要 0.6 美元、缓存读取才 0.015 美元的国产极速模型，结果两边全部一次跑通，真的让我蛮惊喜的， 但是兄弟们，更让我觉得离谱的是左边这个便宜一大截的国产极速战神M3.1 Flash： 代码初始化阶段，它自己检测出 4 个潜在的渲染报错，不等人催，直接在后台跑测试、当场修好，把帧率稳稳焊死在 60 帧。 关键是我的提示词里压根没要求做控制面板啊，它还主动在右上角多搓了一套，轨道线、辉光和星云全都能一键开关，交付出来就是开箱即用的完整作品，这个是真的有点牛逼。 右边的 Opus 5.5 则守住了审美的底子：太阳光晕带着强烈的电影镜头感，底部导航栏和木星、海王星的资料卡排版极其讲究，光影和设计品味上，老牌旗舰Claude依然有它独一份的细腻。 但把两边的账单摊在桌面上，任何开发者都绕不开这道残酷的算术题： 在绝大多数日常前端交互、代码构建和智能体流程里，两者的能力差距，已经远远配不上它们之间的价格鸿沟。 以前咱们搞复杂项目，总觉得必须全程挂着最贵的脑子。 亲手测完这把我才明白，真正能把智能体推向全天候量产的，往往不是高高在上的奢侈品， 而是M3.1 Flash这种选手，它就是前阵子在榜单上匿名跑了 200T Token 的 Space Bunny，亮明真身后速度拉到上一代 M3 的 3 倍、成本砍掉三分之二，靠极低成本的多轮自检，不知疲倦地替你干完脏活累活。 不过惊艳归惊艳，跑分咱们也得客观看：它拿全场第一的是自动化办公 Automation-Bench（62.2），硬核终端编程 Terminal Bench 4.0 只有 38.1，离 Opus 5 的 51.8 还有实打实的代差。 啃十万行祖传代码库，还得请Claude旗舰出山哈哈，但日常七八成的原型开发和自动化脚本，它已经能稳稳接过去，光这一点就够团队省下大把真金白银了hh。 说实话到现在我心里还感慨的不行，国产模型进化迭代的速度真的越来越快了，也越来越强了，当一个每百万输出才 0.6 美元的快模型学会了自己跑测试、自己抓虫修到 60 帧，国外顶配旗舰靠单次智商筑起的溢价围墙，就已经开始漏风松动了。 看完这组 3D 太阳系，更打动你的是左边功能完备、还能自己抓虫的极客风，还是右边更有电影感的设计风？欢迎铁汁们互动交流， 反正要是让我自己掏腰包跑批量任务，左边这种又快又省还不让人操心的选手，我已经挑不出拒绝的理由了哈哈哈。