原文概要

OpenAI 发布了 GPT-6.1 Terra,定位写得很直接:快、省,主攻精细视觉推理、精确的 computer use,以及更清楚自然的写作。它的标准 API 价格比 GPT-5.6 Terra 和 GPT-6.1 Sol 都低 50%,$1 每百万输入 token、$0.05 每百万缓存输入、$5 每百万输出,API 名叫 gpt-6.1-terra;Plus、Pro、Enterprise、Edu 全部套餐可以在 ChatGPT Work 和 Codex 里直接用。HN 上这条帖子拿到 341 分,两百多条评论。

官方给的数字几乎都围绕成本。Agents’ Last Exam 上,medium 推理档比 GPT-5.6 Terra 高 15 个百分点,成本约为三分之一;和 GPT-6 Astra 基本打平,但每 token 价格只有它的十分之一;最高档下 token 效率是 Claude Opus 5.5 的三倍多、价格只有四分之一,折算下来单任务成本低约 9 倍,速度比 GPT-6 Astra 和 GPT-6.1 Sol 都快三倍以上。MMMU-Pro 和 OSWorld 2.1 上,Terra 也都追平或超过上一代,接近 Astra 的水平。

事实性是这代改动最大的一块。在 OpenAI 内部的「诱导错误」prompt 基准上,低推理档的错误率从 GPT-5.6 Terra 的 53.6% 掉到 16.7%,降了近 37 个百分点;最高档则降到 4.3%,接近 GPT-6 Astra 和 GPT-6.1 Sol。代码侧,Terminal-Bench 4.0 在最高推理档解出 50.3% 的任务,是 GPT-5.6 Terra 公布成绩 21.5% 的两倍多;但官方也承认,最复杂的编码任务仍然更适合用 GPT-6.1 Sol 或 GPT-6 Astra。安全方面,Terra 达到了 Preparedness Framework 网络安全部分的 Critical 阈值。

发布之后,评论区迅速发现了一件比发布本身更值得讨论的事:Terra 这个名字本来已经因为 GPT-6.1 Sol 的上市而被停产,如今又以 6.1 的版本回来。产品线始终是 Luna < Terra < Sol < Astra 四档,名字对应的是规模、也是价格——问题是,Sol 6.1 现在在很多任务上比 Astra 还强,而 Terra 6.1 又插在中间。于是「我该选哪个」这个问题,比这次模型本身更热。

讨论焦点

一、评论区真正想要的不是 Terra,是 Opus 5.5

第一反应几乎一致:你们要发的是能打 Opus 5.5 的模型,发下来的却是「便宜版的中间档」。enraged_camel 的说法最短:

“People keep asking OpenAI for a mode that’s as good as Opus 5.5, and they release this instead. Interesting strategy. Let’s see how it works out for them.” — enraged_camel [c:50248259] (译文:大家一直在要求 OpenAI 出一个跟 Opus 5.5 同水平的模式,结果他们发了这个。有意思的策略。看看效果如何吧。)

cbg0 顺着这个逻辑往下推了一步:

“If they had a model on par with Opus 5.5 there would be no need to undercut it by that much, that’s probably the logic.” — cbg0 [c:50248552] (译文:如果他们手上有一个能和 Opus 5.5 掰手腕的模型,就没必要把价格压这么低——大概就是这个逻辑。)

Silagi 则觉得这篇发布里的 benchmark 套 benchmark 已经套到没边了:

“That’s wild to say when comparing to the OpenAI family of models. I’ve never seen so many tests that validate tests that validate tests.” — Silagi [c:50248532] (译文:拿 OpenAI 自家这一族模型横向比,这话实在离谱。我从没见过这么多「验证 benchmark 的 benchmark 再去验证 benchmark」的东西。)

也有更悲观的判断。cloverich 认为 OpenAI 手上根本没有对标 Opus 5.5 的产品,只有一个传闻中的 “Bel”:

“They don’t have anything competitive with Opus 5.5 yet. There is a supposed model in the works, “Bel” last time I paid attention to gossip, that is thought to be better, later this year.” — cloverich [c:50249238] (译文:他们目前还没有任何能和 Opus 5.5 竞争的产品。上次我听八卦时提到有个在做的模型叫 “Bel”,据说更好,大概今年晚些时候出来。)

bananaflag 用 Anthropic 自己的产品线做类比,反过来给 OpenAI 留了个台阶:

“I guess it’s the same as Opus 5.5 being in a way better than any Fable.” — bananaflag [c:50248514] (译文:我猜这跟 Opus 5.5 比任何 Fable 都好上一截是一个道理。)

Archit3ch 希望的下一个别再是成本故事:

“I rather enjoy OpenAI’s focus on cost-per-task. I’m optimistic about them tackling frontier intelligence next. Astra 6.1 please?” — Archit3ch [c:50248461] (译文:我还挺喜欢 OpenAI 这次盯住单任务成本这件事的。我对它们接下来去啃前沿智能乐观。求你了,出个 Astra 6.1 吧?)

二、唯一不好看的那一栏:Terminal-Bench

发布里大部分数字都往好的方向写,只有一个例外,而且被反复揪住。官方原文写的是 Terminal-Bench 4.0 在最高推理档解出 50.3%,对比 GPT-5.6 Terra 的 21.5%。问题在于,同一批评论里说 Opus 5.5 在这个基准上是 70%。

Starlevel004 的怀疑最具体:

“That terminal bench 4.0 result is… bad? Terra XHigh more expensive than Sol Low and still does worse.” — Starlevel004 [c:50248472] (译文:那个 Terminal Bench 4.0 的成绩……很差吗?Terra 开 XHigh 比 Sol 开 Low 还贵,结果还更差。)

“It’s very likely this is a next checkpoint of GPT-6 “Sol”, over-optimised for the web chat and offered to Codex subscribers as a “please go away” offering.” — Starlevel004 [c:50248472] (译文:这多半就是 GPT-6 “Sol” 的下一个 checkpoint,被过度优化到只适合网页聊天,然后作为「求你别用了」的选项端给 Codex 订阅用户。)

user43928 的总结更短:

“Looks great on the first chart, but terrible on the Terminal Bench 4.0 where it needs xhigh to come close to 6.1 Sol, and costs more for a slightly worse result.” — user43928 [c:50248481] (译文:第一张图看着很棒,但在 Terminal Bench 4.0 上很糟——要开到 xhigh 才勉强接近 6.1 Sol,而且更贵、结果还差一点。)

数字上的对照也被人翻了出来。babelfish 查的是 Agents’ Last Exam:

“ALE (top benchmark in post) is 60% for Terra vs 63% for Opus - I guess we’ll have to wait for AA Index!” — babelfish [c:50248470] (译文:ALE(帖子里的头牌基准)Terra 是 60%,Opus 是 63%——我们大概只能等 Artificial Analysis 的指数了!)

enraged_camel 直接反驳了前面那条「四分之一价格就是平手」的说法:

“…no it’s not? At max effort, it scores 50% on terminal bench vs. Opus 5.5’s 70%.” — enraged_camel [c:50248444] (译文:……不是好吗?最高档下它 Terminal Bench 拿 50%,Opus 5.5 是 70%。)

entrope 的问题指向发布方式本身:

“Taking the benchmarks (as reported by OpenAI) at face value, why wouldn’t they show Opus 5.5 on the OS world benchmark, and show no Anthropic models at all on the “factual error rate” benchmark” — entrope [c:50248902] (译文:既然照 OpenAI 自己报的数字看,那为什么不在 OSWorld 基准上放上 Opus 5.5?为什么在「事实错误率」这个基准上一个 Anthropic 模型都不放?)

三、pelican 已经不新鲜了

这代发布照例配了「画一个骑自行车的鹈鹕」这类视觉任务。评论区对这类基准的耐心已经耗尽。vdfs 说得最重:

“IMO it’s useless because we don’t even know what a pelican on a bike is supposed to look like, It’s very impressive that that models can make these directly using SVG, but we are way past the “wow” stage and test that can’t be gamed and are easy to prove how good the model is in that specific arena are hard to find.” — vdfs [c:50250212] (译文:在我看来它没用,因为我们连「骑自行车的鹈鹕」该长什么样都不知道。模型能直接用 SVG 画出来确实很厉害,但我们早就过了「哇」的阶段——那种没法被刷分、又容易证明模型在这个特定领域有多强的测试,本来就难找。)

他还有一句更短的补充:

“pelican is also nitpicked (and probably the most useless)” — vdfs [c:50249317] (译文:鹈鹕这个题目也是挑出来的(而且大概是其中最没用的一个)。)

simianwords 说得更冷:

“The benchmarks are already saturated and the curves for Terra stick to the axes.” — simianwords [c:50248279] (译文:这些基准早就饱和了,Terra 的曲线都贴着坐标轴。)

怀疑的重点不在具体数字,而在选择性呈现。tristanj:

“Not necessarily, they may have selected benchmarks where Terra performs relatively well compared to those models. Pretty much every model release these days just cherry-picks & highlights the benchmarks where their latest model performs best. We’ll have to wait for the complete benchmark suite to have a better picture.” — tristanj [c:50248430] (译文:也未必,他们可能挑了 Terra 相对表现好的那些基准。现在每次模型发布基本上都是挑挑拣拣,只 highlight 自家新模型成绩最好的基准。等完整的基准套件出来,才能有更好的判断。)

最后落到一个原则。mwigdahl 说他根本不信 benchmark:

“No, I have decided I don’t trust benchmarks without using the model first.” — mwigdahl [c:50250128] (译文:不,我已经决定了,不亲手用过模型就不信 benchmark。)

四、「cost per task」是不是话术

「单任务成本」是这次发布的中心词,评论区对它的解读并不客气。mapontosevenths 认为这是套餐用量被砍之后的补救说法:

“They’re doing its because they recently cut plan usage by about 50%, leaving Anthropic providing about 5x the value on their plans.” — mapontosevenths [c:50249789] (译文:他们这么做,是因为最近把套餐用量砍了大约 50%,这让 Anthropic 的套餐看起来提供了约 5 倍的价值。)

“The “cost per task” metric is just a weaselly way to sound like they didn’t just double the cost for most users, and aren’t having their proverbial lunch money stolen. They can no longer claim to provide more tokens, more capable models, or more usage in general. So they made up a new metric where “task” could be defined to mean whatever they need it to mean.” — mapontosevenths [c:50249789] (译文:「单任务成本」这个指标只是换个说法,好让自己听起来没有把大多数用户的成本翻倍、也没有让人吃掉本该属于自己的那份午餐钱。他们已经没法再声称自己提供更多 token、更强的模型或者更多额度,于是就发明了一个新指标——里面的「任务」可以定义成他们需要的任何东西。)

Archit3ch 立刻反驳这个前提:

“5x is calculated in terms of API costs, nobody pays those extreme Anthropic prices outside enterprise users.” — Archit3ch [c:50250233] (译文:5 倍是按 API 成本算的,除了企业用户,没人付得起 Anthropic 那种离谱的价钱。)

“Most users aren’t on their $200 plan. Most users (including on the $200 plan) continue being subsidized.” — Archit3ch [c:50250233] (译文:大多数用户并不在 200 美元那档套餐上。大多数用户(包括 200 美元那档)依然在享受补贴。)

真正在意的指标是另一套。aavaa 说得直接:

“Cost per task, time per task, quality (good luck measuring objectively between models) are the metric that matter. I don’t care how much a token costs, I’m not buying tokens, I’m buying solutions.” — aavaa [c:50250125] (译文:单任务成本、单任务时间、质量(想在模型之间客观比较这个?祝你好运)——这些才是要紧的指标。我不在乎一个 token 多少钱,我不买 token,我买解决方案。)

rfgplk 则替 Anthropic 算了笔账:

“Anthropic models are comically overpriced. They also gobble up an obscene # of tokens to get the same result as others. The price delta is literally 5x.” — rfgplk [c:50248988] (译文:Anthropic 的模型贵得离谱。他们还要消耗多得离谱的 token 量,才能得到和别人一样的结果。价差就是 5 倍。)

五、命名:从一条线变成一个盒子

真正占了最多评论数的,是产品线命名。Someone1234 把混乱的来龙去脉捋了一遍:

“I’m very confused about their model market position/naming schema.” — Someone1234 [c:50248491] (译文:我对他们的模型市场定位/命名体系非常困惑。)

他随后把当时的排位列了出来:Luna(最小)、Terra(中)、Sol(大),这套在大家熟悉了「行星」之后还算顺;后来他们又加上 Astra 作为「更大的那个」,变成 Luna、Terra、Sol、Astra 四档。真正让他看不懂的是接下来的两步:

“But then things go confusing, they released Sol-6.1 which sometimes is superior to Astra, and discontinued Terra which made it look like this:” — Someone1234 [c:50248491] (译文:然后事情就乱了——他们发布了 Sol-6.1,它有时候比 Astra 还强,同时又停产了 Terra,于是看起来就成了这样:)

也就是说,排位变成了 Luna(最小)与 Sol-6.1 等同于 Astra。而现在他们又把 Terra 以 6.1 的名义请回来,让这件事更混乱了:

“Now they’re re-introduced Terra, as a 6.1 release, making this even more confusing. Did Microsoft’s chief branding officer move over to OpenAI?” — Someone1234 [c:50248491] (译文:现在他们又把 Terra 以 6.1 的名义请回来,让这件事更混乱了。微软的品牌负责人是不是跳槽去 OpenAI 了?)

the__alchemist 给出了整段讨论最抽象、也最准的描述——坐标系从一维变成三维:

“Hey! I think it’s confusing because there is a mix of three components: - Model name - Version number - Effort level” — the__alchemist [c:50248903] (译文:我觉得之所以让人困惑,是因为这里混了三个维度:模型名、版本号、effort 档位。)

“All 3 of these are a one-dimensional balance of price/speed, and capability (Loose terms). The confusion is in how these overlap, how they’re announced/marketed etc. So you go from 1 dimensions to 3; you can visualize it as a box.” — the__alchemist [c:50248903] (译文:这三者各自都是一条「价格/速度」与「能力」(宽泛地说)之间的权衡轴。困惑在于它们怎么互相叠加、怎么被发布和营销。所以你就从 1 个维度走进了 3 个维度——你可以把它想象成一个盒子。)

jedberg 提供了另一种读法:名字是价格线,能力是交叉的。

“Part of the confusion is that it’s not linear capabilities. One model is not strictly “better” than the other. They have orthogonal capabilities.” — jedberg [c:50248621] (译文:困惑的一部分在于,能力不是线性的。一个模型并不严格「比」另一个更好。它们的能力是正交的。)

“I think the names are more of a reference to price. Luna is always cheaper than terra which is cheaper than Sol which is cheaper than Astra (for the same model number).” — jedberg [c:50248621] (译文:我认为这些名字更多是价格的指代。在同一个模型版本号下,Luna 永远比 terra 便宜,terra 比 Sol 便宜,Sol 比 Astra 便宜。)

但这条说法随即被指出有矛盾——名字是线性的,能力不是。jedberg 自己补的那句就是:

“That’s why “They have orthogonal capabilities” matters. You need to find the one that’s best suited to your particular task. But the model names are linear by price.” — jedberg [c:50250447] (译文:所以「它们的能力是正交的」才重要——你得找到那个最适合你具体任务的那个。但模型名是按价格线性排的。)

stymaar 把这条缝补得更细:

“Per token, yes. But per task it’s not even necessarily the case depending on how much thinking is needed.” — stymaar [c:50248932] (译文:按 token 算是。但按任务就不一定了,取决于需要想多少。)

Closi 建议干脆别用行星名:

“Or on the flipside, if the models were called GPT 6.1 Mini and GPT 6.1 Max, nobody would be arguing about the names being hard to grok. The words ‘mini’ or ‘small’ are more descriptive than ‘Luna’ or ‘Terra’.” — Closi [c:50250413] (译文:换个角度说,如果模型叫 GPT 6.1 Mini 和 GPT 6.1 Max,就没人会争论这些名字有多难懂了。「mini」或者「small」比「Luna」「Terra」更有描述性。)

torben-friis 把「说不清」具体化成了两个该说清却没说的问题:

“Understanding how effort and class interact (Is a high effort terra above a low effort Sol? Are larger models mostly always better regardless of effort?)” — torben-friis [c:50249434] (译文:effort 档位和模型档位是怎么互相作用的(高 effort 的 terra 比低 effort 的 Sol 强吗?更大的模型是不是在绝大多数 effort 下都更好?))

“What kind of issue warrants going up in effort? what kind of issue warrants going up in model class? What kind of issue is unlikely to be solved by either?” — torben-friis [c:50249434] (译文:什么样的问题值得提高 effort?什么样的问题值得提高模型档位?什么样的问题两者都不指望能解决?)

六、OpenAI 员工下场解释,评论区一半在吐槽

讨论进行到一半,一位自称在 OpenAI 工作的用户 tedsanders 出来说明产品逻辑,这是整条帖子里回复最多的一条评论,下面挂了二十多条回复。他的解释分三层。

第一层是能力多维,第二层是规模,第三层是这次发布的定位:

“In general, LLM performance is incredibly multidimensional. There’s no guarantee of a fixed ordering where the model that is best at math is also the model that’s best at jokes. Despite that complexity, much of the variation in performance is explained by just a single variable: model size.” — tedsanders [c:50249160] (译文:一般来说,LLM 的表现是多维得离谱的。没有任何保证说数学最好的模型也就是讲笑话最好的模型。尽管如此,表现差异的很大一部分其实只由一个变量解释:模型规模。)

“As you note, our models are ordered: Luna < Terra < Sol < Astra” — tedsanders [c:50249160] (译文:正如你说的,我们的模型是有序的:Luna < Terra < Sol < Astra。)

“In the case of GPT-6.1 Terra, I’d say Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things.” — tedsanders [c:50249160] (译文:就 GPT-6.1 Terra 而言,我大概会说 Terra 在技术沟通上比 Astra 和 Sol 更好,在视觉和 computer use 上差不多(但快得多!),其他大多数事情上稍微差一点。)

他收尾那句「如果你对怎么给产品起名有建议,我洗耳恭听」,在评论区被抓住不放:

“For some reason, this is just hilarious to me. Idk if you’re being serious or if this is passive aggressive but the thought of OAI actually taking advice from a random HNer and basing their naming scheme off that is peak.” — halfmatthalfcat [c:50249375] (译文:不知为什么,这让我觉得特别好笑。不知道你是认真的还是阴阳怪气,但想到 OAI 真的会听一个随机 HN 用户的建议、还照着改命名方案,就很离谱。)

不过也有人认为,这条回复本身就是最好的产品说明。GodelNumbering:

“I think the very response you wrote here is a better form of marketing for the software engineers than what openAI puts out there. When everything is groundbreaking and everything tops charts, nothing is.” — GodelNumbering [c:50249650] (译文:我觉得你在这里写的这条回复,比 openAI 官方放出来的那些更像给工程师看的营销文案。当什么东西都是「开创性突破」、什么东西都登顶排行榜时,就什么都不算。)

他的诉求也很具体:把 Terra 不如高阶模型的地方也画出来。

“I suspect a lot of target demographic (developers) prefer simpler and honest communication about not just the models high points but low points too.” — GodelNumbering [c:50249650] (译文:我怀疑目标人群(开发者)里有很多人会更接受简单、诚实的沟通——不只是讲模型的高点,也讲低点。)

他的诉求也很具体:把 Terra 不如高阶模型的地方也画出来。在那篇 Terra 的发布里,如果他能找到一张明确标出「Terra 在哪些用例上比更高档模型更差」的图,整体上会更容易被说服。

atombender 的问题更接近普通用户视角:

“But when you want “maximum intelligence, cost and time be damned”, there’s the implication that bigger models are smarter on a linear scale, and every provider’s model release naming is currently conflicting with that expectation.” — atombender [c:50249604] (译文:但当你要的是「最大智能,成本和时间都不管」的时候,潜台词就是更大的模型在线性尺度上更聪明,而目前各家厂商的模型命名都跟这个预期打架。)

“This isn’t just a naming problem, but a communication problem. If I knew that Astra was the best at coding but a bit bad at writing, and Sol was almost as smart but better at technical writing, then that’s the kind of thing I expect to be documented.” — atombender [c:50249604] (译文:这不是单纯的命名问题,而是沟通问题。如果我知道 Astra 最擅长写代码但写作一般,Sol 聪明程度接近但技术写作更好,那就属于我期待被文档化的事情。)

unethical_ban 直接替 OpenAI 写出了那份缺失的说明书:

“They should have a guide saying “Terra is best at technical documentation and communication, Astra is best at large software projects and research, Sol is good for development” or something like that.” — unethical_ban [c:50250270] (译文:他们应该出一份指南,写上「Terra 最适合技术文档和沟通,Astra 最适合大型软件项目和研究,Sol 适合日常开发」之类的内容。)

而 tedsanders 本人的回答很诚实:他自己也没测过多少。

“I’d imagine Terra might have the most readable prose, but Astra might be more likely to catch bugs or mistakes to highlight.” — tedsanders [c:50250664] (译文:我猜 Terra 的文字可能最好读,但 Astra 可能更容易挑出 bug 或者值得指出的错误。)

“I wouldn’t want to make a recommendation without doing a lot of testing, and I haven’t done a lot of testing.” — tedsanders [c:50250664] (译文:不做大量测试的话,我不想给推荐,而我并没有做大量测试。)

七、那到底该怎么选

最实际的一派建议是:别猜,自己跑。elxr 说得最干脆:

“Run your own evals then if you have this much of a need to “correctly” choose. That’ll make it clear.” — elxr [c:50249948] (译文:那你跑自己的评测啊,如果你这么需要「正确」地选一个的话。那样就清楚了。)

nailer 不接受这个甩锅:

“No. OpenAI want my money, it’s on them to communicate properly.” — nailer [c:50250103] (译文:不。OpenAI 想拿我的钱,把话说明白是他们的责任。)

mediaman 的反问最接近「你们到底要什么」:

“But what’s the ask? This is the cleanest way I see of them doing it. I see a ton of criticism and not a single person coming up with anything better, unless it’s to take away options, like removing effort level.” — mediaman [c:50249164] (译文:那到底要什么?这已经是我能想到的最干净的做法了。我看到一大堆批评,却没看到一个人提出更好的方案——除非是把选项拿走,比如取消 effort 档位。)

“Why are engineers using this as a professional tool expecting this to be bought like a consumer gadget and then blaming the maker?” — mediaman [c:50249164] (译文:为什么工程师把它当专业工具用,却指望它像消费电子一样买,然后转头怪厂商?)

brianpan 指出这套复杂度首先坑的是普通用户:

“These names are not just for engineers. Why is it ok for engineers to have confusing names and ways to interact with something?” — brianpan [c:50249685] (译文:这些名字不只是给工程师的。为什么可以让工程师面对一堆难懂的名字和难懂的交互方式?)

工具层面其实已经有答案,只是没人指向它。andai 提到两个现成的曲线来源:

“Oh, easy: https://deepswe.datacurve.ai/ Artificial Analysis has their own curves which are pretty similar.” — andai [c:50250291] (译文:简单:https://deepswe.datacurve.ai/ Artificial Analysis 自己也有曲线,样子差不多。)

ilaksh 补了半句现实提醒:

“You can start with Artificial Analysis benchmarks but, then you need to test in the real tasks anyway.” — ilaksh [c:50250025] (译文:你可以先看 Artificial Analysis 的基准,但最后还是得在真实任务上测。)

至于「换个版本要不要跟着换」,margalabargala 的问题最实在:

“If I build something around this Terra 6.1 release, then when Luna 6.2 is released in a month, should I swap?” — margalabargala [c:50249599] (译文:如果我基于这次 Terra 6.1 的发布做了产品,那一个月后 Luna 6.2 发布时,我该不该换?)

为什么不一次把四个模型全发了,也是同一条线上的追问:

“What prevents OAI (and Ant for that matter) from releasing new version for all models at once? Instead, we got Opus 5.5, Sonnet 5.5 and Haiku 5.5 released weeks apart, and so are Sol 6.1 and Terra 6.1?” — mi_lk [c:50249823] (译文:是什么阻止了 OAI(Anthropic 也一样)一次性把所有模型的新版本都发出来?结果 Opus 5.5、Sonnet 5.5、Haiku 5.5 隔了几周才发,Sol 6.1 和 Terra 6.1 也是这样?)

RussianCow 给了一个纯商业的解释:

“My guess is nothing, and they’re currently just releasing them individually as soon they’re ready instead of all at once so that they hit the market more quickly. AI moves quickly enough that a delay of a couple weeks could result in significant lost revenue.” — RussianCow [c:50250313] (译文:我猜什么也没有,他们只是现在选择一个准备好了就单独发,而不是等齐了一起发,好更快打进市场。AI 变化太快,拖上几周可能就是相当可观的收入损失。)

八、编程之外的另一半

还有一批评论者不太关心命名,只关心这次发布对写代码之外的事有没有帮助。Imustaskforhelp 写了一大段,其中最见性情的是这几句:

“Honestly I’d much rather take this model then. It is having improvements in other fields than coding while being more cost efficient.” — Imustaskforhelp [c:50250527] (译文:老实说,那我更宁愿要这个模型。它在编程之外的领域有提升,同时成本效率也更高。)

同一个人也说了这次发布让他意识到的事:

“Maybe this is the reason why we have gotten almost an exponential growth in coding capabilities as that became the most (if not only) capability that truly really matters.” — Imustaskforhelp [c:50250527] (译文:也许这就是为什么编程能力几乎在指数增长——因为编程成了唯一真正要紧的能力(如果不是唯一的话)。)

“As an (aspiring) developer (currently in 1st year college), I don’t wish for AI to replace coding even more for me or to push it even more or to suck the joy out of it even further.” — Imustaskforhelp [c:50250527] (译文:作为一个(有志成为)开发者(目前在大学一年级),我不希望 AI 进一步替代我写代码,也不希望它把这件事推得更远、更从我身上抽走乐趣。)

wyre 的判断很直接——这个模型的受众不是写代码的人:

“I don’t think software devs are the audience for 6.1 Terra, but instead computer work, automation, and technical writing.” — wyre [c:50249711] (译文:我不认为软件开发者是 6.1 Terra 的目标用户,它面向的是 computer work、自动化和技术写作。)

onlyrealcuzzo 补上了 Terra 回归的产品逻辑:Sol 6.1 质量上去了,但慢到没法用,除非按 Astra 的价格买:

“It’s nice that they have Terra which seems to fill a void that’s been needed since Sol 6 focused on cost (at the expense of quality), Sol 6.1 got a huge boost in quality, but became unusably slow unless you paid Astra prices - at which point, why not just use Astra? Now at least we have a viable other option again.” — onlyrealcuzzo [c:50249120] (译文:他们有 Terra 挺好的,看起来补上了一个从 Sol 6 专注成本(牺牲质量)以来就一直存在的空缺:Sol 6.1 质量大幅提升,但慢到没法用,除非你按 Astra 的价格买——那为什么不直接用 Astra?现在至少又有了一个可行的选项。)

tamimio 则贡献了一个使用细节:

“Right?! I think sol has some obsessive tendencies, i literally caught it the other day spawning a sub sub agent doing a final fourth test for a piece of code that isn’t even critical say in embedded or similar, now I have to instruct it to only test once!” — tamimio [c:50250545] (译文:对吧?!我觉得 sol 有某种强迫倾向,我前几天就抓到它为一个根本不关键的代码片段(嵌入式之类的)又派生了一个子子 agent 去做第四次最终测试,现在我不得不专门告诉它只测一次!)

典型观点一览

立场 用户 一句话
想要的是 Opus 5.5 平替 enraged_camel 你们发的不是那个
压价幅度本身说明问题 cbg0 有平手就不必砍这么狠
benchmark 套 benchmark Silagi 从没见过这么多层验证
OpenAI 还没有对标 Opus 5.5 的产品 cloverich 传闻中的 Bel 还没出
Terminal-Bench 是唯一难看的数字 Starlevel004 XHigh 更贵还更差
ALE 差 3 个百分点 babelfish 得等 AA Index
鹈鹕基准已经过时 vdfs 早就过了「哇」的阶段
benchmark 只挑自己赢的 tristanj 等完整套件出来
不亲手用就不信 benchmark mwigdahl 我不信任数字
「单任务成本」是话术 mapontosevenths 套餐用量刚砍了一半
5 倍是按 API 成本算的 Archit3ch 普通用户还在被补贴
我买的是解决方案不是 token aavaa 单任务成本/时间/质量
Anthropic 贵得离谱 rfgplk 价差就是 5 倍
命名体系已经失控 Someone1234 微软品牌负责人跳槽了?
一维变三维 the__alchemist 名字、版本号、effort
名字是价格线,能力是正交的 jedberg 名字线性,能力交叉
换成 Mini/Max 就没这些争议 Closi 「mini」比「Terra」有描述性
两个该说清而没说清的问题 torben-friis effort 与档位如何交互
员工下场解释 tedsanders 规模解释了大部分差异
官方回复比官方博客更像营销 GodelNumbering 什么都登顶就什么都不算
该出一份模型能力指南 unethical_ban Terra 写文档,Astra 做研究
跑自己的评测 elxr 别猜
沟通是厂商的责任 nailer 他们想拿我的钱
到底要什么 mediaman 想拿掉选项请直说
名字不只是给工程师的 brianpan 普通用户怎么办
曲线图其实早有了 andai deepswe 与 Artificial Analysis
换版本要不要跟 margalabargala 一个月后 Luna 6.2 来了怎么办
为什么不一次全发 mi_lk 隔几周才发一个
不是为开发者准备的 wyre computer work 与技术写作
编程能力的指数增长 Imustaskforhelp 编程成了唯一要紧的事

总体情绪

这场发布最特别的地方,是官方给的每一个数字都在往「买」的方向推,而评论区几乎所有高赞内容都在往「怎么买」的方向问。

一边是数字的漂亮:Agents’ Last Exam 上和 GPT-6 Astra 打平、每 token 只要它十分之一的价格,单任务成本比 Claude Opus 5.5 低约 9 倍,速度快三倍以上,事实错误率从 53.6% 掉到 16.7%。这些数字没有一条是假的——按 OpenAI 自己的口径。另一边是随之而来的那个问题:四档模型、effort 档位、版本号,一个模型该选哪一个。OpenAI 自己给出的答案是「跑自己的评测」,而这恰好是付费用户最不想做、也最难做的事。

最有意思的裂缝出现在 tedsanders 那条回复里。他用四段话把体系讲清楚了,评论区一半在吐槽「OpenAI 竟然来问我们该怎么给产品命名」,另一半在向他索要那张不存在的矩阵图。而他自己承认没做多少测试,等于把整场讨论又送回起点:Terra 到底比 Sol 强在哪,没人知道,厂商似乎也不知道。

这大概是史上最省事的一次发布——新模型的价格算清楚了,而「挑哪个」的责任,连同那张缺失的表格,一起交给了用户。

引用帖子

# 标题 URL
1 Introducing GPT-6.1 Terra(原文博客) https://openai.com/index/introducing-gpt-6-1-terra/
2 HN 讨论 https://news.ycombinator.com/item?id=50248124
3 GPT-6.1 Terra system card https://deploymentsafety.openai.com/gpt-6-1-terra
4 Anthropic 订阅提供 5 倍价值(评论中引用) https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x
5 deepswe 曲线(评论中提到) https://deepswe.datacurve.ai/
6 Artificial Analysis(评论中提到) https://artificialanalysis.ai/

免责声明

本文为 Hacker News 讨论的中文摘要,所有引文归原作者所有,本文仅作整理。讨论涉及 OpenAI 新模型 GPT-6.1 Terra 的性能与定价宣称、模型命名体系、产品线发布节奏,以及「单任务成本」这类营销口径的争议。文中引用的性能与价格数字——Agents' Last Exam、MMMU-Pro、OSWorld 2.1、Terminal-Bench 4.0、Terminal-Bench-Science 0.1 的成绩,53.6%/16.7%/4.3% 的错误率,约 9 倍的单任务成本差,$1/$0.05/$5 的每百万 token 价格,5 倍套餐价值——均转述自 OpenAI 官方发布与 HN 读者在帖内的说法,未经独立核实,请以原始链接为准。引文均带有 `[c:id]` 形式的评论编号,可在原帖通过 `item?id={id}#{c_id}` 形式定位。读者应自行判断各方立场与解读。

本摘要由 AI 模型辅助生成:minimax-cn-coding-plan/MiniMax-M3