ChatGPT Dot and Grok Bot: One is included in any package priced starting at $100, while the other has just had its price lowered to $20.

On August 11, SpaceXAI and Cursor launched Grok Bot. On September 29, OpenAI released dots on DevDay. On October 30, ChatGPT Pro’s quota was first announced to be cut in half and then doubled within a month. On November 8, dots learned to get the verification code from your email. On November 9, Grok Bot’s threshold dropped from $200 to $20 per month.

In three months, the two "resident agents" have changed their rules five times. If we don't clarify this timeline clearly, today's comparison of who is more cost-effective is basically comparing two expired price lists.

Both do the same thing: give you a computer that is turned on in the cloud, you close your laptop, and it will click on web pages, fill out forms, and run scheduled tasks for you. The difference is all in the details.

Whoever runs out first will step into the trap first

Grok Bot was in early beta the day it was launched, and was only available for the three most expensive subscriptions: SuperGrok Heavy for $300 per month, Cursor Ultra for $200 per month, and Cursor Teams Premium for $120 per seat per month. It will be released to SuperGrok Plus, Cursor Pro+ and all Teams ten days later on August 21st. As of this week, Cursor Pro ($20) and SuperGrok ($30) are also available.

dots is the other way around. It will be sold together with Pro 100, Pro 200, Pro 500 and Business Premium when it comes up, and the first dot will be given for each package. Free, Go and Plus are not available.

So the answer to the question "Which one is cheaper?" has changed once in three months.

What does that computer look like down there?

The description of Grok Bot is relatively realistic: each account has a resident Linux virtual machine with a browser, terminal and a /workspace directory. The specifications compiled by the evaluation are Debian, 128GB storage, and 16GB memory. The browser login status and files will always be retained.

Dots are each dot with its own cloud computer and independent browser. Access to your local computer is closed by default. The cloud browser does not inherit your local login, so the first time you ask it to order a restaurant, it has to log in again. Before November 8, when encountering a SMS or email verification code, OpenAI's help page allowed you to take over the browser yourself; now you can authorize it to read the verification code yourself from the connected mailbox, with two options: one-time or permanent.

Sharing a computer for all Bots is a big problem

This is most easily overlooked.

Each bot you create under Grok Bot shares the same computer, the same set of files, the same browser session and command line credentials. xAI itself states in the documentation: Do not regard different Bots as security boundaries. If you delete a Bot, the shared files and browser logins it leaves behind will not be deleted. You have to manually go through six steps to clean them up.

Give a specific scenario. You log in to the company's CRM for a Bot to sort out leads, and create another Bot to go to several forums every day to capture information. The latter Bot reads a webpage with hidden instructions. It can encounter not only the forum, but also the CRM that has been logged in.

The isolation of dots is more complete, but the price is that each dot has to log in to your account again, which makes it more cumbersome to use. The newly added verification code reading function in November is equivalent to moving one step closer to saving trouble between "security" and "trouble saving". If you click "Always allow", when a dot is led astray, it will hold all the accounts that can receive verification codes in your email address.

Neither is perfect, but the risks of Grok Bot are there by default, while the risks of dots are set by you.

The parts not printed on the price list

$20 to get access to a Grok Bot sounds delicious. The available amount is reset on a weekly basis, and the excess amount will be deducted from the account according to the original price of the model and token. There is currently no spending limit that can be set.

During the beta period, some users said on Hacker News that the biggest drawback was token consumption, which increased dramatically in a month. Another person bought Cursor Ultra specifically to try it out. As a result, the backend panel showed zero usage, but 48% had been used in the app. The most unsettling aspect of this type of product is that the bills don't match up.

The problem with dots is that the numbers are not released at all. How much does the second dot cost, how much does it cost to upgrade to speed up and increase the workload, and whether background autonomous tasks are charged on a pay-per-use basis? OpenAI has not written yet. Chatting with dot does not count towards the ChatGPT quota, but the work it does in Codex or ChatGPT Work will be counted into the quota of those two products.

This brings us back to that repetition. At the end of September, OpenAI announced that starting from October 30, the Codex and Work quota of Pro 200 will be reduced from 20 times to 10 times of Plus, and GPT-6 Pro will be reduced from 200 to 100 per week. A post was opened in the community at that time, with a title to the effect of "Pro 200 is pretty good now, please stop optimizing." As a result, on October 30, OpenAI’s Tibo announced that the usage limits of Pro 100, Pro 200, Pro 500 and Business Premium would all be doubled, and the dot quota would be calculated separately and retained as usual.

Two adjustments in opposite directions in one month. If you order Pro now, it’s okay to calculate the amount based on the current amount, but don’t expect it to still be the same amount in six months.

Looks like it's overturned

There are three sentences in the Grok Bot documentation that are worth copying and pasting on the monitor: the test run will "perform real operations"; approval only blocks actions that have not been done yet, and cannot undo those that have been completed; and clicking "Stop Now" will not undo anything. That said, there's no drill mode. You let it try out a routine of "sending follow-up emails to customers in batches", and the emails will actually be sent.

Platform coverage also has pitfalls. The desktop supports macOS, Windows, and iPhones above iOS 18. Linux, Android, and iPad are not supported at launch. There is also a "Download for Linux" button on the marketing page. The GitHub login for some iOS users is directly broken and cannot be accessed. The audit log that the team version said to be done says "coming soon" twice in the document. Currently, you can only view the usage panel and the chat records of each Bot. None of the compliance certifications such as SOC 2 and ISO 27001 are stated.

The problem with dots is concentrated on "restriction". It must be created in a desktop app or computer browser, and mobile web pages cannot be built; it cannot actively call you when it is first launched; Pro users in several European regions simply cannot use it. There are not many early reviews, and they are relatively scattered. It is not yet a conclusive reputation.

I don’t quite agree with the common saying on the Internet

The common conclusion is: the bottom layer of dots is GPT-6 Astra, and the model is stronger, so it is more valuable.

When doing resident tasks, the strength of the model is often not the bottleneck. What is really stuck is the website blocking automation, popping up the verification code, and asking to log in again. Both companies can only transfer these steps to others. Grok Bot officials themselves admit that some processes "require more care." A routine that summarizes emails and calendars for you every day can be completed with a weaker model; a process that requires logging into ten backends, no matter how strong the model is, will still get stuck on the third verification code.

Three things determine the experience: how to save the login status, whether it can be rolled back if an error occurs, and whether the bill can be capped. On these three fronts, both companies are now only halfway there.

put it on yourself

For developers who are already using Cursor Pro to write code, Grok Bot is now giving away a trial seat for free. Try opening a Bot that only does low-risk things, and don’t log in to any accounts involving money or customer data.

For those who have already subscribed to ChatGPT Pro, the first dot is already included in the package, so there is no reason not to use it; the option to read the verification code is "only this time".

If you haven’t subscribed to either of them and just paid for this feature, you’ll have to wait another month.

The pricing and quotas of the two companies have not yet stabilized, and Grok does not even have a spending limit.

Related comparisons: Meta Muse, Beanbag Work, OpenClaw: Three agent routes, who will tell you the difference? · Claude Ultra and ChatGPT Pro 200: The 5-hour window is gone, only one writes how many times the weekly quota is · Aside, Browser Use, ChatGPT Browser Agent: 99% running score and a real migration, three things are missing in the middle