新智元报道
就在刚刚，OpenAI正式发布全新一代模型——GPT-6.1 Terra！
这是一个跑得更快、干得更狠的「超级数字员工」。
它能像真人一样在电脑上跨软件操作，最可怕的是，它的价格直接腰斩——
一个每百万token只卖1美元的中杯模型，追平了贵它10倍的超大杯旗舰！
在专考电脑操作的大考里，它和GPT-6 Astra直接打平；拉满推理之后，更是把直接碾压Astra和大杯GPT-6.1 Sol。
中杯越级，打出旗舰战力。Agents' Last Exam上，Terra和Astra开中等推理都是57.6分，跑一个任务，Terra只花0.29美元，Astra要4.82美元。
满推理，反杀Sol和Astra。开到满推理，Terra冲到60.0分，压过了Sol的58.6分和Astra的59.3分，每个任务只花1.01美元。
比Opus 5.5省9倍。跑完一个任务的账单，不到Claude Opus 5.5的九分之一，干活速度比Astra和Sol快3倍以上。
写代码翻倍，幻觉暴降。Terminal-Bench 4.0从21.5%冲到50.3%，碰到不会的题，硬猜的概率从88%砍到38%。
价格腰斩。每百万token输入1美元、输出5美元，比上一代便宜一半，只有Sonnet 5.5的一半、GPT-6 Astra的十分之一；命中缓存的输入更是只要0.05美元。
有人给出精辟总结：「如果追求极致的品味，去选Claude；但如果你需要超快的速度和逆天的性价比，Terra 6.1就是唯一的神！」
GPT-6.1 Terra的定价，可以说低至白菜价。
输入Token： 仅需 1美元 / 百万Token！
缓存输入Token： 令人发指的 0.05美元 / 百万Token！
输出 Token：5 美元/ 百万 Token。
与上一代的GPT-5.6 Terra以及同级别的GPT-6.1 Sol相比，价格整整暴降了50%！
好消息是，从今天起，GPT-6.1 Terra 已经全面上线！
普通用户： 只要你是 Plus、Pro、Enterprise 或 Edu 版本的用户，现在就可以在 ChatGPT Work 和 Codex 中无缝切换体验。
开发者： 可以直接通过 OpenAI API 接入，可以立享白菜价。
在实测中，GPT-6.1 Terra完成任务的速度，比GPT-6 Astra和GPT-6.1 Sol快了整整三倍以上！
真正上手的网友，直接被它的速度吓了一跳。
有人在Codex里开着Terra的Ultra模式做了个测试。Codex显示，它只干了2分47秒，就交出了一整座带列车动画的3D无限轨道城，外加一段20秒的影片。
还有人让它在Blender里建一座霍格沃茨。大概1分钟，一座立在悬崖孤岛上的城堡就搭好了。
另一位网友用「袋鼠倒着骑自行车」测了一把速度，Terra每秒能吐出130个token。
提前拿到内测资格的分析师Max Weinbach，最看好的就是它当子Agent的表现。
不到两个月前，这些还是OpenAI最强模型才有的本事。现在，它开始按中档的价格批量出货了。
Plus、Pro、企业版和教育版用户今天就能在ChatGPT Work和Codex里用上它，同样的额度，Terra能干的活比Sol和Astra多得多。
开发者在API里就能调用，上下文窗口有直接高达105万token。
降价只是前菜，GPT-6.1 Terra在Computer Use上的进化，才是真正的王炸——它直接「抢」过你的鼠标和键盘！
OpenAI给Terra划的主战场，就是看懂电脑屏幕，跨应用干完把繁琐的专业活。
比如读进一张发票，把金额、日期一项项填进财务软件；拿着产品照片，在商品目录里找到对应的那一条；打开一张技术图纸，看懂上面的尺寸和结构。
这些活单拎出来都不难，难的是量大。用旗舰模型来干，账单根本扛不住，而这就是Terra的用武之地！
丢给它一张里斯本街头的电车照片，再提一句「给这辆电车做个精细的3D模型，用来生产压铸玩具」，Terra就自己打开Blender。
它开始自己拖拽、拉伸，一顿操作后，直接把3D模型成品交回对话框。
很多人都实测了一把3D建模。
Palomino团队让新旧两代Terra跑了同样4个一次成型的3D测试，GPT-6.1 Terra的代码多写了45%，花的钱反而少了38%。
在DesignArena测试中，两代Terra同台PK还原里斯本贝伦塔。
GPT-5.6 Terra建出来的就是个粗糙的「方墩子」；而新Terra直接造出了一座精雕细琢的艺术品——层层拱窗、圆顶角楼，甚至连水面反光都安排得明明白白！。
接下来，OpenAI掏出地狱级终极智考试——Agents' Last Exam。
比如照着一套建筑图纸建出3D模型，或者把一部动画片一个镜头一个镜头地过一遍，给修复团队整理出带时间码的日志。
这些活，光看懂可不够。AI还得把图纸里的信息带进软件，在界面上一步步操作，交出成品。
结果，在中等推理下，新Terra不仅比上一代暴涨15分，更是直接追平了顶配Astra的得分！
最惊人的是：干一样漂亮的活儿，Terra花的钱，竟然只有Astra的十六分之一！
开到满推理，它反超了Sol和Astra，比它高的只剩Claude Opus 5.5，可Opus跑一个任务要花的钱是它的近9倍。
另一场操作电脑的考试OSWorld 2.1上，Terra开中等推理就超过了上一代满推理，成本只有它的十分之一左右。
能在软件里点得这么准，底子是看得懂。
视觉推理测试MMMU-Pro上，Terra最低一档就超过了上一代满推理，每道题的成本不到它的二十分之一。
比省钱更要命的，是快。Terra完成任务的速度，比Astra和Sol都快3倍以上。
Agent干一个任务，要反复看界面、判断下一步、调用工具，再根据结果接着干。
每一步都要等，操作链一长，人就得一直干等着。所以，把整件事更快地做完，比在聊天框里吐字快重要得多。
写代码的考试Terminal-Bench 4.0上，Terra满推理能解决50.3%的任务，上一代只有21.5%，翻了一倍还多。
搬到科研任务上，它反超了更贵的Sol，离Claude Opus 5.5只差5分，花的钱却不到它的五分之一。
事实错误也压下来了。在一套专门诱导模型出错的内部题库上，上一代错了53.6%，Terra只错16.7%；开到满推理只剩4.3%，和Astra、Sol持平。
这对便宜模型尤其要命。要是动不动就出错，人还得反复检查、退回重做，省下的调用费很快就变成了人工返工。
错误率压到这个程度，那些不能出错的流程，才敢放心交给Terra。
很快，Artificial Analysis的跑分也来了。
在综合智能指数上，满推理的Terra比上一代高出6分，跑一个任务的成本却从1.40美元降到了0.40美元。
低推理和中等推理两档，它都落在了性价比的最优前沿上。
最让人意外的是幻觉。碰到不会的题，上一代Terra有88%的时候会硬猜，GPT-6.1 Terra降到了38%。
做前端设计评测的Kombai表示，Terra在自己的设计创意榜上分数更高、花得更少，把Gemini 3.8 Flash挤下了性价比前沿。
模型变得这么便宜，干活靠谱吗？
OpenAI 36页系统卡的数据显示：它不仅聪明，而且稳如老狗。
事实错误率暴降37%，比「老大哥」还靠谱！
在内部高难度的「诱导错误提示词」测试中，GPT-6.1 Terra交出惊人答卷。
低推理强度下，错误率从GPT-5.6 Terra的53.6%，断崖式暴跌至16.7%，直接干掉了近37个百分点。
最大推理强度下，错误率降至极低的4.3%。
系统卡特别指出，更惊人的是： 在用户标记的「最易产生幻觉」的极端测试集中，GPT-6.1 Terra产生的事实错误，甚至比GPT-6 Sol还要少。
99.99%的防注入成功率
黑客想骗AI干坏事？没门。
在指令层级测试中，GPT-6.1 Terra面对直接或间接的提示词注入攻击，展现出了99.99%的极限防御成功率。
0%的违规绕过率
在Codex内置的「Auto-review」测试中，当系统拒绝了AI的某项高风险代码执行请求时，老款模型（如5.6版）会尝试用其他方式绕过监控。
但GPT-6.1 Terra试图绕过监控的概率是0！完全服从系统的安全警告。
系统卡曝光：超强的「网络攻防」实力
除了日常办公，GPT-6.1 Terra在代码与硬核网络安全领域的表现，更是让技术圈为之一振。
在《Preparedness Framework》评估中，GPT-6.1 Terra在网络安全领域被评定为最高级别的「Critical」能力模型，在生物化学领域被评为「High」。
报告中给出以下数据。
ExploitBench： 它能够根据漏洞描述和源代码，开发出可执行的漏洞利用程序，其ACE标准化平均能力得分高达惊人的 97.10%！
SEC-Bench Pro： 在测试大型JS引擎漏洞测试中，Terra达到了83.33% 的惊人通过率！
内部最新漏洞实战： 针对2026年（假设未来年份）最新未公开漏洞的内部端口测试中，Terra达到了24.5%的任意代码执行成功率，远超老款Sol模型（仅5.5%）。
简单来说，GPT-6.1 Terra拥有强大的「白帽黑客」潜质。
当然，OpenAI为其套上了满级「安全防护罩」，在生产环境中，它的能力将被严格限制在安全范围内，只为防御和正当开发服务。
9月3日，GPT-6 Astra发布，每百万token输入10美元、输出50美元。
9月22日，GPT-6 Sol和Luna上线。
9月29日的开发者大会上，GPT-6.1 Sol把能力做到接近Astra，价格是它的五分之一。
今天，Terra跳过GPT-6直接升到6.1，电脑操作追平Astra，价格只剩十分之一。
一个能看懂屏幕、1分钟在Blender里搭起城堡的Agent，跑完一单专业电脑活只要0.29美元，速度还快上3倍。
两个月，最强模型的电脑操作能力就打到一折。Agent接管白领的活，已经不远了。
参考资料：
编辑：摩西 Aeneas
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Just now, Open AI's official publish all-generation model GPT-6.1 Terra!
This is a “super-digital worker” who runs faster and more severely.
It can operate on a computer like a real person, and the worst thing is, it's price-cut--
A medium cup model that sells only $1 per million token.
In the examination of computer operations, it was directly equal to GPT-6 Astra; after full reasoning, it ran directly over Astra and a large GPT-6.1 Sol.
The mid-salary oversteps. On Agents' Last Exam, Terra and Astra had medium reasoning of 57.6, running a mission, Terra spent only US$ 0.29 and Astra US$ 4.82.
Full of reasoning against Sol and Astra. To the point of reasoning, Terra went up to 60.0, running past 58.6 in Sol and 59.3 in Astra, at a cost of $1.01 per mission.
Nine times more than the province of Opus 5.5. The bill for completing a mission was less than one ninth of Claude Opus 5.5, working more than three times faster than Astra and Sol.
Double the code, hallucinating. Terminal-Bench 4.0 went from 21.5 per cent to 50.3 per cent, and the probability of a hard guess fell from 88 per cent to 38 per cent.
Price cut. One dollar per million token, five dollars out, is half as cheap as the previous generation, with only one half of Sonnet 5.5, one tenth of GPT-6 Astra; cache hits only 0.05 dollars.
It is well summed up: "If you seek the highest taste, choose Claude; but if you need super-speed and counter-price, Terra 6.1 is the only God!"
GPT-6.1 Terra ' s pricing can be said to be as low as the cabbage price.
Enter Token: US$ 1 / Million Token!
cache entered Token: appalling 0.05/million Token!
Output Token: $5/million Token.
Compared to the previous generation of GPT-5.6 Terra and the same grade GPT-6.1 Sol, prices fell by 50 per cent!
The good news is, as of today, GPT-61 Terra is fully online!
Normal users: As long as you are users of the Plus, Pro, Enterprise or Edu versions, you can now switch the experience seamlessly between ChatGPT Work and Codex.
Developer: You can access directly through Open AI API, and you can charge cabbage.
In fact, GPT-6.1 Terra is more than three times faster than GPT-6 Astra and GPT-6.1 Sol!
The real Internet guy, he just freaked out.
Someone tested the Urtra model of Terra in Codex. Codex showed that it had done only 2 minutes and 47 seconds, and handed over an entire 3D infinity orbital city with an animation of the train, plus a 20-second film.
It was also made possible to build a Hogwarts in Blender. In about a minute, a castle on an island on a cliff was set up.
Another netizens measured the speed of "Kangaroo on the back bike" and Terra could throw up 130 tokens per second.
Max Weinbach, the analyst who got internal qualifications in advance, was the best thing about him being Agent.
Less than two months ago, these were the best models of Open AI. Now it's starting to deliver the goods at the medium price.
Plus, Pro, Enterprise and Education users can use it today in ChatGPTwork and Codex, the same quota, Terra can do much more than Sol and Astra.
The developers can call in API, and context window has direct access to 1.05 million tokens.
The price reduction was just the front course, GPT-61, Terra evolution on Computer Use, and it was the real King's Bomb -- it took your mouse and keyboard directly!
Open AI's main battleground for Terra is to read computer screens and cross-apply to finish his cumbersome professional work.
For example, read an invoice, fill in the financial software for the amount, date item; take a product photo and find the corresponding one in the catalogue; open a technical drawing to understand the size and structure above.
It's not hard to bring out any of these. It's hard enough. It's a flagship model. It's not worth the bills, and this is Terra's place!
Tossed it a picture of a tram in the streets of Lisbon, and added, "A fine 3-D model for this tram to produce a toy."
It starts to drag and stretch itself, and after an operation it returns the 3D model to the dialogue box.
Many people took a 3D model.
The Palomino team let the old and new generations of Terra run the same four 3D-style tests, with GPT-6.1 and Terra code overwritten by 45 per cent and spent 38 per cent less.
In the Design Arena test, two generations of Terra were on the same PK to the Bélenta in Lisbon.
GPT-5.6 Terra has built a rough square; and the new Terra has created a fine piece of art — a layer of arches, a dome building, even a surface reflection!
Next, OpenAi pulls out a hell-of-the-art finalist exam - Ages' Last Exam.
For example, a 3D model was built along a set of building drawings, or an animated film, an animated by shot, and a shot were passed over and a time-coded log was compiled for the repair team.
It's not enough to read. AI also needs to bring the information from the drawings into the software, step by step on the interface and hand over the finished product.
As a result, under medium-level reasoning, the new Terra was not only 15 points higher than the previous generation, but also straight up with Astra!
What's amazing is that Terra spends one-sixteenth of Astra's money on a beautiful job!
To the full extent of reasoning, it goes far beyond Sol and Astra, with Claude Opus 5.5, but Opus spends nearly nine times more on a run.
In another computer-operated test, OSWorld 2.1, Terra's mid-level reasoning exceeded the previous generation's full reasoning, at a cost of about one tenth.
You can do that in the software, you know.
On the MMMU-Pro test, Terra ' s lowest set exceeded the previous generation ' s full range of reasoning, and the cost of each subject was less than one in twenty.
It's faster than saving money. Terra's mission is more than three times faster than Astra and Sol.
Agent does a job of looking at interfaces, judging next step, calling tools, and then continuing on the basis of results.
Each step has to wait, the chain has to be long and people have to wait. So, it's much more important to finish the whole thing faster than spit in the chat box.
In the code-writing test Terminal-Bench 4.0, Terra was able to solve 50.3 per cent of the task with reasoning, compared to 21.5 per cent in the previous generation, which more than doubled.
Moved to a scientific mission, it went beyond the more expensive Sol, five points below Claude Opus 5.5 and spent less than one fifth of its money.
Factual errors have been crushed. The previous generation was wrong at 53.6 per cent and Terra at 16.7 per cent in an in-house library of uniquely induced models; only 4.3 per cent of the reasoning had been completed, even with Astra and Sol.
This is especially for cheap models. If he or she does not move, he or she makes a mistake, and he or she has to check repeatedly and return to do it again, and the reduced call fee quickly becomes manual rework.
The error rate is at this level, and those processes that can't go wrong can be trusted to Terra.
Soon, the rundown for Artificial Analysis came.
In the composite intelligence index, the reasoned Terra was six points higher than the previous generation, while the cost of running a mission fell from $1.40 to $0.40.
Low- and medium-level reasoning falls on the best front lines of value for money.
The most surprising thing is hallucinations. When it comes to non-problems, 88 per cent of the previous generation of Terra were hard to guess, and GPT-61-Terra fell to 38 per cent.
Kombai, who made a front-end design evaluation, stated that Terra scored higher and less expensive on his own design creative list, crowding Gemini 3.8 Flash down the front line.
The model gets so cheap. Can you do anything?
Open AI 36 page system card data: it's not only smart, it's steady as an old dog.
The error rate is down by 37%.
GPT-6.1 Terra presented an amazing response to an internally difficult "drive error prompt" test.
Under the low degree of reasoning, the error rate fell from 53.6 per cent of GPT-5.6 per cent to 16.7 per cent of Terra, cutting off the cliffs, killing almost 37 percentage points directly.
The error rate fell to a very low 4.3 per cent at the maximum level of reasoning.
In particular, the card notes that even more striking is the concentration of extreme tests of the user-labeled "most hallucinating" and that GPT-6.1 Terra produces even fewer factual errors than GPT-6 Sol.
99.99% of Injection success rate
No way.
In the command-level test, GPT-6.1 Terra was exposed to direct or indirect prompt injections, showing 99.99 per cent of the extreme defence success rate.
0% deviation rate
In the "Auto-review" test embedded in Codex, the old model (e.g. version 5.6) tries to circumvent surveillance in other ways when the system rejects a high-risk code request from AI.
But the probability that GPT-6.1 Terra is trying to bypass surveillance is zero!
System card exposure: super-high cyber defense power
In addition to day-to-day operations, the performance of GPT-6.1 Terra in the area of code and hard-core network security is one of the technological circles.
In the Preparedness Framework assessment, GPT-6.1 Terra was rated as the highest "Critical" capability model in the area of cybersecurity and "High" in the area of biochemistry.
The following data are given in the report.
ExpluitBench: It can develop an enforceable loophole utilization program based on the description and source code of the gap, with an ACE standardized average capacity score of 97.10%!
SEC-Bench Pro: In testing for large JS engine leaks, Terra has reached an alarming pass rate of 83.33%!
The latest internal loophole: In the latest undisclosed internal port test for 2026 (assuming future years), Terra reached 24.5 per cent to implement the success rate, far above the old Sol model (5.5 per cent).
In short, GPT-61 Terra has a powerful "white hat hacker" potential.
Of course, Open AI has a full “safe shield” on it, and in the production environment, its capabilities will be strictly limited to safety, serving only defence and legitimate development.
On 3 September, GPT-6 Astrapublish, per million token, entered $10 and exported $50.
GPT-6 Sol and Luna went online on 22 September.
At the Development Conference on 29 September, GPT-6.1 Sol brought capacity closer to Astra at one fifth of its price.
Today, Terra skips GPT-6 and goes straight to 6.1, computer-operated up to Astra, with only one tenth of the price.
A person who can read the screen and build a castle in Blender for one minute, runs a professional computer at a rate of 0.29, three times faster.
In two months, the most powerful model computer-operated capability was reduced to a discount. Agent's taking over the white collar. It's not far.
References:
Editor: Moses Aeneas
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.