新智元报道
把活交给AI，派哪个AI上场也归AI管了。
11月2日，OpenAI的ChatGPT和Codex负责人Tibo在X上连发两条推文，剧透了dots的一次关键升级：
同一项任务，dots将拆成几段，交给不同模型接力。
在Tibo第一条推文里，他透露驱动dots的模型阵容正在扩容：旗舰GPT-6 Astra之外，Terra即将入列，后面还会塞进更多型号的模型。
真正重要的改变，藏在他的第二条推文里：
你的dot会为任务挑选最合适的模型，在工作的不同部分调用不同模型。
别小看这句话，它改变了AI干活的逻辑。
以前用AI，是选定一个模型干到底；升级后的dots，要一事一议、按步换模型。
对用户来说，这个变动很关键。
那些执行的步骤可以交给速度快的模型，你就不用干等；只把一些烧脑的核心环节交给旗舰模型，也能把好钢用在刀刃上。
这样，你就能让AI在把活干好的同时，把时间和成本都打下来。
dots是OpenAI的个人智能体，9月29日在DevDay上发布。
它自带云端电脑，全天候在线，能盯着你的目标连轴跑长时任务。
首发时，给它当大脑的是当家旗舰GPT-6 Astra。
Astra已经够聪明，为什么还要塞一个Terra？
Tibo的答案是：Terra更快，「性格」也升了级（聊起来也更讨喜）。
先看实力。
第三方评测平台Vals.ai的综合榜上，Terra得分60.17%。在28个参评模型里，它排第10。
虽然跑分不算拔尖，但它精确命中了用户最在乎的两大痛点：快，而且便宜。
OpenAI官方还甩出一个数字：
在电脑操作评测中，Terra完成任务的速度，是Astra和Sol的3倍以上。
但这个数字只是单项成绩，如果放进dots实战，整套活干下来，大概率快不了3倍。
因为评测比的是单科百米冲刺，实际工作是一场障碍接力赛。
一项任务总耗时多久，除了要看最难啃的骨头，也要看模型之间的交接是否高效。
与Astra相比，Terra的强项是电脑操作。
看懂屏幕、填表格、点按钮，在不同软件之间来回切换……它能抠出来的时间，主要就在这些动手不动脑的环节。
可真正烧脑的那一步，该想多久，还得想多久，要留给旗舰模型。
比起速度差距，价格相差则更为悬殊。
按标准API的普通输入、输出单价，Astra均为Terra的10倍，Sol均为Terra的2倍。
虽然，API单价不等于dots的订阅费，也不等于你办成一件事的总账单。
可积少成多。
dots是个一跑就是几十步的长时任务智能体，单步省一点时间、一点钱，乘上几十个回合，省的就不是一星半点。
让旗舰模型负责「镇场子」，专攻最难的推理任务就够了，如果让它包揽每一步就等于浪费算力和时间。
过去很长一段时间，挑模型是用户自己的苦差事。
这件事，连Tibo自己都忍不了。
今年10月，Tibo在Lenny’s Podcast谈到使用AI的繁琐设置时，就吐槽：
你简直得有个「模型选择器博士学位」。
Tibo谈模型选择和推理设置给用户带来的额外负担。
想让AI干活，你得先摸清每个模型的脾气，再估一估这步活有多难。
还没开工，前面就有一道超难的选择题。
所以dots从诞生第一天起，就砍掉了模型选择器。
你唯一要选的，是在哪个渠道跟它搭话。
这次升级，把选择的颗粒度再往深推一层：从按任务选模型，细化到按步骤换模型。
举个职场人最熟悉的例子。
你让dot做一份项目延期汇报，整个流程要走四步：
先从文档里扒出进度和预算；再核对几张表格里的数据；然后分析项目为什么延期；最后把结论排成PPT。
这四步，难度也完全不在一个档次上。
扒资料，拼的是眼疾手快；表格数据打架时，判断哪个口径可信最吃推理；做PPT，又考验提炼、排版和动手操作。
如果全程都让最贵的Astra包办，就像请首席科学家帮你填Excel，你的钱包和耐心会一起滴血。
可如果全交给快模型，到了分析延期原因这一报告的关键环节，又容易想不透或者答偏。
最聪明的解法是流水线作业：哪一步谁擅长，就让谁上。
目前，官方还没亮出具体调度规则，不能就此断定分析一定归Astra、排版一定归Terra。
这是同一个dot在后台换模型干活，和一群智能体同时开工是两码事。
因此对用户来说，前台也将极其清爽：你只管提要求，后台派谁上场dot自己排班。
替用户选模型这件事，OpenAI在一年多前就干过，还翻过车。
2025年8月7日，GPT-5发布。
当时的王牌卖点是实时路由，每条提问由谁回答，系统自动分配。
发布前，奥特曼还特意吐槽过旧版模型菜单，嫌它乱成一团。
没想到打脸来得飞快，上线第一天，路由就趴窝了。
奥特曼不得不公开认栽：路由一度停摆，GPT-5因此显得笨了不少。
折腾几天后，OpenAI不得不把模型菜单加了回来，GPT-4o也重新回到付费用户的列表里。
一年多过去，一些用户的心理阴影还在。
中文技术社区LINUX DO上，已经有网友警觉：这会不会又是「动态降智」？
也有人在Tibo推文底下追问：那dots现在用的到底是Sol吗？
模型选择权一旦收进黑盒，用户可能连谁在替自己干活都分不清。
更何况，这次的调度更难。
GPT-5的路由，只替每一次提问挑模型，是一锤子买卖。dots要在同一项长任务里排班，玩的是多棒接力。
排班越细，交接出错的风险就越大。
前一步扒出的资料，换个模型能不能完整接住？中途改了一个数据口径，做PPT的模型会不会还在用旧数据？某一步跑偏了，系统能不能当场揪出来？
长链条任务里，单步跑得再快只是表面，全程不返工才是里子。
交接一旦翻车，快模型省下的那点时间，还不够填返工的坑。你是愿意花钱看AI表演，还是想让它们真正把活干好？
替用户做选择，真正难的地方是省下算力，还能不把活干砸。
Tibo在推文中说：Terra会帮助OpenAI，把dots带给更多人。
现在的dots，还是高端玩家的专属。
个人用户想尝鲜，得先订每月至少100美元的Pro。20美元的Plus用户，至今还用不上。
还有一位网友说得很实在：
Tibo，我不想要dot，我想要Plus多给点用量。
如果dots足够便宜呢？他还想不想要？
官方谈的是功能升级，用户惦记的还是老额度。
更快、更便宜的Terra，让dots走出「100美元俱乐部」又更近了一步。
长期追踪AI新功能的TestingCatalog，看到Tibo的预告只回了一句：有待实测。
效果怎么样，还要过上线后的四个实测检验：
完成一项复杂任务，到底要等多久？
跑一次要烧掉多少额度？
中途要人工救火几次？
最终交付的东西，能不能直接交差？
对于用户来说，这四个问题比模型列表里多一个Terra重要得多。
但风向已经变了：模型选择器没有消失，只是从你的屏幕上转到了智能体的大脑中。
如今，Meta有Muse，马斯克有Grok Bot，Google有Gemini Spark，都在抢「个人第一助手」的入口。
用户把任务交出去，想省下的不只是那几分钟的打字时间，还有挑模型、等结果、改错重来的全部心力。
在这个战场上，拼的早就不止旗舰跑分，而是谁能用最少的等待和折腾，交出一份不用返工的活。
最聪明的模型已经不再稀缺，会看活派人、一次把活干利索的管家，才稀缺。
参考资料：
编辑：元宇
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.
Give it to AI, which one is in AI's hands.
On November 2, ChatGPT of Open AI and Tibo, head of Codex, issued two tweets on X, which were a key upgrade of dots:
In the same task, dots will be split into sections and handed over to different models for relay.
In Tibo's first tweet, he revealed that the model that drives Dots is positive. In addition to amplification: the flagship GPT-6 Astra, Terra is about to enter and more models will be inserted.
The real change is hidden in his second tweet:
Your dot will select the most appropriate model for the mission, using different models in different parts of the work.
Don't underestimate that. It changes the logic of AI working.
Previously, AI was used to select a model to dry it out; upgraded dots, with a step-by-step model.
This change is critical for users.
Those steps can be delivered to fast-paced models, so you don't have to wait; just hand over the core of the brain burn to the flagship model and use the steel on the blade.
That way, you can get AI to take time and cost while working.
Dots are Openai's personal AI agent, on DevDay on 29 September.
It carries its own cloud-based computer, it's all-weather online, and it can keep an eye on your target axis for a long time.
When it started, it was the flagship GPT-6 Astra.
Astra's smart enough. Why do you need a Terra?
Tibo's answer is that Terra is faster and that "genre" is promoted (and even more pleasurable).
Let's start with strength.
Terra scored 60.17 per cent on the consolidated list of the third-party evaluation platform Vals.ai. Of the 28 evaluation models, it ranks 10.
Although running points are not high, they strike exactly two of the pain points that users care about most: fast and cheap.
Open AI also lost a number:
In the computer operations assessment, Terra completed its mission at a rate more than three times that of Astra and Sol.
However, this figure is a single achievement, and it would probably be less than three times faster if it were put into the dots and the whole job was done.
Since the evaluation was a single-combranium sprint, the actual job was an obstacle relay.
The total length of a mission depends on the efficiency of the inter-model interface, in addition to the most difficult bones.
In comparison to Astra, Terra's strength is computer operations.
Read screens, fill tables, dot buttons, switch back and forth between different software... It's the only time it's gonna come out, mostly in the middle of it.
But the real brain burn, how long to think, how long to think, left to the flagship model.
Price differentials are even greater than the speed gap.
Astra is 10 times as much as Terra and Sol is twice as much as Terra for standard API input and output unit prices.
Although the API unit price is not equal to the subscriptions to dots, it is not equal to the total bill of one thing.
It's a small amount.
Dots are a few dozen steps away from the long mission of AI agent, one step away, a little money, dozens of rounds, not a star and a half.
If the flagship model were to be responsible for the town field, it would be enough to focus on the most difficult reasoning tasks, and if it were to cover every step, it would be a waste of effort and time.
For a long time, the choice of the model was the user's own hard work.
Even Tibo can't stand this.
In October of this year, when Tibo spoke at Lenny's Podcast about using the cumbersome settings of AI:
You have to have a Ph.D. Model Selector.
Tibo talks about the additional burden placed on users by model selection and reasoning settings.
To get AI to work, you have to figure out each model's temper and figure out how hard it is.
We haven't started yet. There's an extremely difficult choice.
So dots cut off the model selector from the first day.
All you have to choose is which channel you want to talk to.
This upgrade places the selected particle size further down: from task-based model to step-by-step model.
One of the most familiar examples in the world of work.
You get dot to do a project extension report, and the whole process goes four steps:
First extract the progress and budget from the file; then check the data in a few tables; then analyse why the project was postponed; and finally rank the conclusions as PPT.
These four steps, too, are completely off the grid.
Quick-eyed, hard-to-earned data; hard-to-read reasoning when table data fights; PPT to test extraction, layout and hands-on operations.
If you let the most expensive Astra fit, it's like asking the chief scientist to fill in Excel, your wallet and patience will drop blood together.
However, if all were to be given to the fast-track model, it would be difficult to understand or respond to the key point of the report, which is analysing the reasons for the extension.
The smartest solution is running water lines: whoever is good at any step, lets go.
At the moment, the official rules for the movement have not yet been shown, and it cannot be concluded that the analysis must belong to Astra and the layout to Terra.
This is a two-way street between the same dot working on the backstage and working at the same time as a group of AI agent.
So, for users, the front desk will also be extremely refreshing: just ask, and who's going to dot's own schedule?
Open AI did it more than a year ago and turned over the car.
GPT-5publish, 7 August 2025.
At that time, the ace was a real-time route, each question was answered and the system was automatically distributed.
Before publish, Ottman threw himself through the old model menu and said it was a mess.
I didn't know the face came in fast. The first day on the line, the router fell asleep.
Otman had to admit publicly that the route had been cut off and GPT-5 had become a bit stupid.
A few days later, Open AI had to add the model menu back and GPT-4o was back on the list of paid users.
More than a year later, the psychological shadow of some users remained.
The Chinese tech community, LINUX DO, already has online alerts: Is this a "dynamic dementia" again?
And there are people who are asking under Tibo: "Is this Sol that dots are using now?"
Once model options are included in the black box, the user may not even know who is working for himself.
Besides, it's harder to move this time.
The GPT-5 route, which singles out models for each question, is a hammer trade. Dots have to line up in the same long assignment, playing with a sticky relay.
The more detailed the shift, the greater the risk of error.
Can a new model be perfected? A data calibration has been changed in the middle. Is it possible to model a PPT with old data? One of the steps is off. Can you find it on the spot?
In a long chain mission, a one-way run is a surface, and it's Lee who does not go back to work.
Once the switch is turned over, the fast model saves the time and is not enough to fill the back-to-work pit. Are you willing to pay for the AI show or do you want them to really do the job?
The real difficulty in choosing for users is to save money and not to blow it up.
Tibo tweets: Terra will help Open AI and bring dots to more people.
Now dots, still high-end player exclusive.
Individual users need to book at least $100 a month for Pro for taste. Twenty dollars for Plus users has not been used to date.
And one of the netizens was quite honest:
Tibo, I don't want dot, I want Plus to give more.
What if dots are cheap enough? Does he want it?
The official talk is about functional upgrading, and users are thinking about old quota.
The faster and cheaper Terra brought Dots out of a $100 club.
TestingCatalog, a long-term tracker of AI's new functionality, saw Tibo's prognosis only say one thing: it's still to be measured.
How's it going?
How long will it take to complete a complex task?
How many quotas do you have to burn once?
How many times do I have to do this?
Can you just hand over the final delivery?
For users, these four questions are much more important than one Terra in the model list.
But the wind has changed: the model chooser didn't disappear, it just moved from your screen to the brain of AI agent.
Today, Meta has Muse, Muse, Mask, Grok Bot, Google, Gemini Spark, and is robbing the entrance of the Personal First Assistant.
Users give up the task and want to save more than just a few minutes of typing time, but also the full effort to pick out models, wait for results, and change things wrong.
On this battlefield, it has long been more than a flagship score, but who can hand over a job without returning to work with the least wait and manoeuvre.
The smartest models are no longer scarce, and it is the human beings and the housekeepers who do the work at once that are scarce.
References:
Editor: Won-woo
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.