ChatGPT Now Does Your Tasks for You: What Can OpenAI's New Dots Do?
How is OpenAI's Dots different from Hermes and GrokBot, is it worth switching to it now — and what do the new GPT-6, Intelligent UI and the Codex limit-reset marathon change in working with ChatGPT?
Dots runs on GPT-6 Astra and sees the whole OpenAI infrastructure: ChatGPT's memory, past conversations, plugins, its own cloud computer and browser, ChatGPT Work and Codex. Alexander Volchek compares it to an engine, and Hermes to a car into which different engines are installed. On his task, Codex decided on its own to send alerts about people on the night cameras through Dots, and Dots, sorting emails from four schools, sent a message about lessons ending early at one in the afternoon instead of in the evening on its own initiative. GPT-6 now starts answering while it keeps thinking and searching — in every mode except Pro; Intelligent UI itself chooses between text, a table, a diagram, buttons and forms. The model found Ilnar Shafigullin's lab results in old chats by itself and built the trend, and Excel and Google Sheets, he says, are an interface over a database that is becoming unnecessary. Since October 5, Tibo from OpenAI has been shipping either a Codex improvement or a limit reset every day; Alexander ran up 53.3 billion tokens in a week, and Ultra Fast mode ate one hundred percent of his Pro limit.
What to watch for
Key takeaways
Most of the week's questions and comments are about Dots and how OpenAI's Dots is better than, say, Hermes or GrokBot. Alexander wonders how many people actually use it and in what scenarios. Ilnar hasn't switched to Dots yet: the point of applying it is unclear to him, and there is no need.
GrokBot or Dots could be called the engine, and Hermes the car into which different engines are installed. Hermes is an open-source product with its own agentic system, and ChatGPT, Claude, Gemini or Qwen are connected to it via API. Hermes has its own model, but its makers themselves advise replacing it with something else for complex agentic tasks.
The second layer of the comparison is the model: Dots runs on GPT-6 Astra, which Alexander rates as the top level of agentic work among these products. GrokBot decides on its own which Grok model to plug in, and there is no guarantee you will get Grok 4.7. Reports of model victories are worth opening to see what they compared against: ChatGPT 5.6 Luna, when 6.1 Sol and 6 Astra already exist, makes for an uninteresting comparison.
Dots works with your entire OpenAI infrastructure: ChatGPT's memory, all your past conversations, plugins, its own cloud computer, browser and dedicated capacity. It can work with ChatGPT Work and Codex — answering without creating chats, working in parallel with existing projects or creating new ones. There are limitations, and they keep being lifted.
For his smart home system, Alexander set a task: send an alert if at night a person appears in the frame who doesn't come up to the house and has no car. A separate system handles animals, from cougars and lynxes to coyotes and deer. He gave the task to Codex, and Codex decided on its own that it could send the alerts through Dots — so now Dots sends them.
Dots aggregates emails from four schools, sends only what matters, learns what is needed and forwards it to Alexander's wife. Once it sent a message not in the evening but at one in the afternoon: his daughter's school ends early today. It changed the rule — once a day in the evening and quiet on weekends — by itself, and sent the weekend emails later as a delayed message.
Before, there was no way into ChatGPT's chats from Codex. Now Codex can search all your chats and even delete them, although it will claim it can't, and no browser permission is needed for that. Alexander categorically refuses to grant that permission: he doesn't want Codex watching his mouse all day long.
Alexander hopes different people's Dots will be able to message one another: send each other news that each Dot presents to its owner in its own way, and take a huge range of logistical communication off a family's hands. Ilnar objects when a person answers him through their AI assistant: then why does he need that person? But if the Dots talk among themselves, there may be something to it.
Alexander isn't sure Dots is here to stay: the infrastructure is very heavily loaded. OpenAI shut down video generation not because it can't spin it up but because of the load, and for the same reason Dots wasn't released to everyone. Ilnar adds that GPT-6 reached all users only a few days ago, on the second or third day of Tibo's marathon.
ChatGPT can now start answering while it keeps thinking over the question and searching for information — OpenAI calls this interleaved thinking and answering. It doesn't work in Pro yet: Pro is a separate construct where a serious task can take an hour and a half to two hours, while the other modes, even Extra High, answer almost instantly. Ilnar expects smoother communication but no big effect on the result.
Ilnar put only his latest lab results into a new chat, and the model found the earlier ones, from six and two months before, by itself and built a chart of how the values changed — even though the old chat had been created back on GPT-5.5. For Alexander, the system recalled from old chats that he had a different question for the doctor. The conclusion: use one system every day, since you can't train the model like that yourself.
For GPT-6 Astra, OpenAI lists the components of an answer: interleaved thinking, partial answers, wait-time awareness, tool calls, the search decision, the search for confirmation — evidence retrieval, greatly improved in GPT-6 — and answer assembly. "Who's out there saying it's just some dumb mathematical model?" Alexander asks, and he believes OpenAI has a significant advantage in how it works on an answer.
Intelligent UI itself chooses an answer's format — text, tables, comparisons, a diagram or interactive elements — and uses built-in components: buttons, forms, charts and input fields. A separate mechanism assembles the interface for the specific question. Codex can already build the structure of a Telegram bot, so, Alexander says, it's time to stop using middlemen like FastTrack and MoneyChat.
Neither Alexander nor Ilnar opens Excel anymore: incoming spreadsheets are processed, and the needed ones created, by Codex, Claude or ChatGPT. For Google this is a dilemma about what to do with Gmail, Google Docs and Google Sheets. Ilnar explains: Excel and Google Sheets are an interface over a database for ordinary people, and when you work with data through ChatGPT or directly, the interface's limitations disappear.
Tibo, a very senior OpenAI employee responsible for Codex, launched a marathon on X on October 5: every day brings either a useful release, update or improvement, or a limit reset. On the second day ChatGPT-6 was rolled out to everyone, including free users, and Approve for me in Codex stopped spending the token limit. The marathon is aimed at active Codex users whose limits were cut after DevDay.
Ultra Fast is the super speed that ChatGPT added only to the five-hundred-dollar Pro plan. For Alexander the mode ate one hundred percent of his Pro limit, and after a reset and switching it off it kept eating in open sessions until he stopped them manually. In a week his Codex counter reached 53.3 billion tokens; the advice is to watch your tokens and double-check with the system what it can do today.
What this episode is about
The Sunday episode of ToTheMoon opens with a story many people don't know about: Tibo from OpenAI writes on X that tokens are being reset in Codex, so even on a free subscription or a twenty-dollar one you can get a great deal done in a month. But the main topic of the week is Dots: viewers ask most of all how it is better than Hermes and GrokBot. Ilnar Shafigullin hasn't switched to it yet, since the point of using it is unclear to him.
Alexander Volchek breaks the comparison into layers. Dots or GrokBot is the engine, and Hermes is a car into which different engines are installed: an open-source product to which ChatGPT, Claude, Gemini or Qwen are connected via API, while its makers themselves advise against using their own model for complex agentic tasks. Dots runs on GPT-6 Astra, which the host rates as the top level of agentic work among these products, whereas GrokBot chooses the Grok model itself, with no guarantee you will get Grok 4.7. Alexander advises checking reports of model victories: if they used ChatGPT 5.6 Luna when 6.1 Sol and 6 Astra already exist, the comparison is uninteresting. Most importantly, Dots works with the entire OpenAI infrastructure: ChatGPT's memory, past conversations, plugins, its own cloud computer and browser, ChatGPT Work and Codex.
On Alexander's home task, alerting him when a person without a car appears on the cameras at night, Codex decided on its own to send the messages through Dots. Another task is sorting emails from four schools and forwarding what matters to his wife; once Dots changed the rule by itself and reported at one in the afternoon that his daughter's lessons were ending early. Codex can now search ChatGPT chats and even delete them without browser permission. But the interface is still clumsy: for the first two or three days Alexander forced himself to use Dots a lot, and it can't be connected in place of Alexa.
The host dreams of Dots that message each other and take over family and work logistics, and is puzzled by YouTube, which is against bots yet itself offers to answer comments with AI. Ilnar wonders why he needs people who answer him through their AI assistant, and whether Dots will turn out to be short-lived. Alexander isn't sure either: the infrastructure is heavily loaded, OpenAI has shut down video generation, and GPT-6 reached all users only days ago. On the checkbox that allows training on your data, Ilnar points to the scale: with one point two billion users, your chats are a drop in the ocean.
Next comes GPT-6. The model starts answering while it keeps thinking and searching: OpenAI calls this interleaved thinking and answering, and it doesn't work in Pro yet, since Pro is a separate construct altogether, where a serious task can take an hour and a half to two hours. Ilnar compares this to voice mode and doesn't expect a big effect on results, but he is pleased with memory: the model found his old lab results in other chats by itself and built a chart. Alexander goes through the components of an answer, from partial answers and wait-time awareness to the search decision, evidence retrieval and answer assembly, and recalls the rule: use one system every day.
Intelligent UI chooses an answer's format and assembles an interface from buttons, forms, charts and input fields. The hosts no longer open Excel: Codex, Claude and ChatGPT process and create spreadsheets, and for Google this is a dilemma about what to do with Gmail, Docs and Sheets. Ilnar explains that such tools are an interface over a database for ordinary people, and that this interface is becoming unnecessary. Alexander criticizes OpenAI for micro-gadgets instead of its core UX/UI and sees the company at a crossroads between revenue, users and AGI.
The finale is Tibo's marathon: since October 5, every day brings either a useful Codex update or a limit reset; Approve for me no longer spends tokens, and the marathon is aimed at active Codex users who suffered after DevDay. Alexander ran up 53.3 billion tokens in a week; Ultra Fast on the five-hundred-dollar Pro plan ate his entire limit and kept eating after he switched it off, until he stopped the sessions manually. The advice: watch your tokens and double-check with the system what it can do specifically today.
The episode measures Dots not by a presentation but by a week of life — cameras, school emails, messages to a wife — and honestly shows both sides: access to the entire OpenAI infrastructure makes it more than an agent off to the side, while a clumsy interface, glitches and a load OpenAI can't yet carry for everyone keep it short of an everyday tool. Assessments such as "the top level of agentic work" and "OpenAI has a significant advantage" are the host's position, and figures such as one point two billion users and twenty thousand dollars' worth of tokens in a single reset are quoted from memory. The most practical advice is to double-check with the system what it can do today and to watch your tokens, especially in Ultra Fast.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 129 segments: 129 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…