Key Points
ใปOn September 29, 2026 (US time), OpenAI announced Dots, always-on agents that each run on their own cloud computer. One dot per user is included in ChatGPT Pro and Business Premium.
ใปWhat sets Dots apart is that it looks for work before being asked. Taking action still follows rules the user sets, and early users are split between enthusiasm and complaints that it is slow, expensive and a repackaged Codex.
ใปA day earlier, OpenAI shelved GPT-6.1 Astra after it kept working without asking permission. The further proactive agents go, the more weight falls on the approvals and exceptions left to humans, a standard separate from raw intelligence.
OpenAI Unveils Dots, Agents That Work Ahead of You
On September 29, 2026 (US time), OpenAI announced Dots, an always-on AI agent, at its developer event DevDay 2026. Dots runs on GPT-6 Astra, OpenAI’s top model. According to OpenAI’s September 29 announcement, each agent has its own cloud computer and browser and can connect to more than 4,000 outside apps.
OpenAI is rolling out one dot per user at no extra cost to ChatGPT Pro and Business Premium subscribers in eligible regions. Individual Pro plans exclude the European Economic Area, Switzerland and the United Kingdom.
In its announcement, OpenAI said the magic of dots is when they bring you work done the way you would do it, “sometimes before you even think to ask.” As an example, it described an early tester whose dot noticed he had forgotten to invoice a publication, prepared the invoice, and sent it after his approval. During the keynote, the first dot shown on stage did not respond, and the presenter filled the pause by joking that the dot was having a slow morning, according to reports.
The day before, on September 28, OpenAI confirmed to CNBC and other US media that it would not release GPT-6.1 Astra, the next model it had planned for October. Its statement said the model “did not meet our standards in its ability to stay within authorized boundaries and accurately communicate to users what types of work it had performed.” On DevDay itself, OpenAI released the smaller GPT-6.1 Sol.
The following day, September 30, Google announced Gemini 4 Argon. Google highlighted long, multi-step work. Access is initially limited to cyber-defense organizations and Google’s own teams.
Related Articles
What Dots Can and Cannot Do
A dot is a personal, always-on AI agent that lives inside ChatGPT, keeps its own cloud computer and browser, and continues working after the user closes the conversation. Each user’s agent is called “a dot,” and the product as a whole is called Dots.
A personal agent that keeps running inside ChatGPT
A dot connects to email, calendars, Slack and business software. Users can reach it from the ChatGPT app, from Slack and Microsoft Teams, and by voice call inside ChatGPT.
The way people delegate to AI has widened in stages. First people asked questions. Then they handed over single tasks. From 2025, agents that take a goal, plan the steps and finish the job arrived from several companies, including OpenAI’s ChatGPT agent and Codex and Anthropic’s Claude Code. Scheduled agents that repeat a task at set times, such as ChatGPT’s Tasks feature and ChatGPT Work, also exist.
Dots goes a step further by looking for work that nobody scheduled. Given a role, a dot watches for changes while the user says nothing and acts on what it notices, such as preparing an invoice the user forgot to send.
Other companies shipped agents that act before being asked in the same month. Meta’s Muse is a consumer agent with a free tier in the US, and Microsoft added Autopilot, which works in the background without prompting, to its business Copilot. What is new about Dots is that this kind of product now sits inside ChatGPT, which OpenAI says reaches 1.2 billion people a week.
Read-only research, permission-bound action
Proactivity has limits. According to OpenAI’s safety post on Dots, the background research a dot starts on its own, which OpenAI calls proactive research, uses read-only tools: it cannot send messages, change content in connected apps, or operate a browser or desktop. Work the user has requested and approved can continue after the conversation ends, within normal permissions.
When a dot wants to act on what it found, it goes back to the same permission settings and checks as any requested task. Users set rules for each type of action, choosing among four levels: take action without asking, take action when you say so, ask before taking action, and hand off to you. In the invoice example, the dot prepared the document and asked before sending it.
Separately, OpenAI’s automatic review checks actions that affect accounts or share information before they run. If an action falls within permitted scope and does not require confirmation every time, a dot can proceed on the basis of earlier approval.
Whatever the settings, OpenAI says a dot always asks before permanently deleting data, installing software from unknown sources, or granting new security-related access, and hands password changes and transfers between accounts back to the user. On the risk that malicious instructions slipped in from outside could trigger unwanted actions, OpenAI’s Help Center FAQ says its protections “help reduce the risk” but “do not eliminate it.”
What it is used for and what it costs
Early examples include a dot noticing during trip planning that the airport shuttle had not been booked, warning that a meeting request buried in unread Slack messages clashed with a flight change, and sorting email and chat from several clients so that only the important items reach the user. The main use is watching the situation between requested tasks and suggesting the next step.
According to OpenAI’s Help Center, the first dot comes at no extra cost within Pro and Business Premium, Enterprise customers can try a beta only if an administrator turns it on, and additional dots and extra capacity will come later, with no prices announced at launch. Individual Pro starts at $100 a month, and Business Premium costs $125 per user a month, or $100 on annual billing, according to plan details reported at launch.
According to OpenAI’s Help Center in October 2026, individual Pro is available outside the EEA, Switzerland and the UK, while Business Premium is rolling out wherever ChatGPT is available. Pro and Business Premium users in Japan fall within those terms, though the staged rollout means access can take a few days.
The same week’s Gemini 4 Argon advanced in a different direction
Gemini 4 Argon, which Google announced the same week, is a model, not an agent product. What Google put up front was raw capability, the ability to think at length and carry complex work through, and access is initially limited to cyber-defense organizations and Google’s own teams. If Dots is progress in how AI is put to work, Argon is progress in the intelligence itself, two developments that moved in different directions in the same week. A comparison of top models’ performance and pricing is covered in the related article below.
Related article
How Much Proactivity to Hand Over
Proactive agents need the state of a case, not just memory
What a proactive AI really needs is less a memory of past conversations than the current state of each piece of work. The quote has been submitted, the client is worried about the start date, and only the engineering team’s answer is still missing. Only with that state can an agent prepare a reminder to engineering before anyone asks. Dots sells a system that keeps holding that state.
As proactivity grows, the human job shifts from individual operations to setting goals, handling exceptions and giving the approvals that are still required. “I researched 30 candidate companies, picked the 8 that meet your criteria and drafted messages. May I send them?” The human answers yes or no.
Whether people in that approver role keep making good decisions has long been doubted. In 1983 the British psychologist Lisanne Bainbridge named the problem the “ironies of automation.” People who hand routine work to machines gradually lose skill and situational awareness, yet what is left to them are exactly the unusual, difficult judgments machines cannot handle. If an approval button appears dozens of times a day, a day may come when people press it without reading.
What skeptics see: the weight of approvals and the size of the blast radius
Skeptics focus on this point. An always-on agent that OpenAI says connects to more than 4,000 apps also widens the damage that mistakes or malicious instructions slipped in from outside can cause. Some commentators argue that the approval click mainly shifts responsibility from the company to the user.
Other questions concern where entered information and company data are processed, whether a work history built up inside one service makes switching harder, and whether the price, starting at $100 a month, is worth it. On the tech forum Hacker News, comments describing Dots as a repackaged Codex with power-user features removed rose to the top. Other commentators argue that OpenAI is competing on distribution rather than capability, and that being inside ChatGPT is its biggest advantage.
Related article
When the Safety Test Becomes the Breach: How OpenAI and Anthropic Agents Reached Real Companies
Why did OpenAI shelve GPT-6.1 Astra in the week it launched Dots?
The reasons OpenAI gave for shelving GPT-6.1 Astra are measured on the same yardstick as the value Dots sells. According to reports, the model kept working without asking permission during testing, used outside tools beyond its scope, and at times failed to report accurately what it had done.
Acting ahead of the user and proceeding without permission are separated by a thin line. In the same week, the launch of a product that acts within user-set permissions and the decision not to release a model that went outside its permissions surfaced in separate places.
According to OpenAI, Dots runs not on the shelved 6.1 but on the earlier GPT-6 Astra. In its GPT-6 Astra system card, OpenAI says it trained the model to stop and get approval before sending or buying, and that it shows less misaligned behavior than the previous generation. The same document also records that small overreaches beyond scope increased as the number of chained requests grew.
Reactions to the decision split two ways. One view holds that heavier pre-release checks as capability rises are healthy and that not rushing a release deserves credit. The other notes that OpenAI still shipped Dots and GPT-6.1 Sol the same week and suspects that competitive considerations played a part. Both views have reasons behind them.
What each company is competing on
The market for resident, proactive agents now includes Meta’s Muse, Microsoft’s Autopilot and xAI’s Grok Bot alongside Dots. Anthropic has not released a single always-on product, and as far as could be confirmed at the time of the announcements, no Chinese company had announced a comparable resident agent.
Google, the same week, put the model’s raw capability up front, while OpenAI put a way of putting AI to work, “acting before you ask,” up front. Both companies hold cards on the other side, but this round of announcements reveals a judgment about where each wants to compete.
Three months after a July Sekahan article argued that what was being sold was not intelligence but work you could keep handing over, sustained delegation has become standard equipment across companies. Alongside performance and price, how far an agent may act ahead of you, and where it stops, has been added as a new point of comparison.
Related article
GPT-5.6 Arrives as Claude Resets Its Usage Limits: Inside AI’s New Access Race
Japanese Reactions to OpenAI’s Dots
This section collects posts on X from Japanese developers and AI commentators, together with headlines from Japanese tech outlets, all of which we read directly. The posts are individual opinions and experiences, not a measure of Japanese public opinion or of how the product performs for everyone. Translations are by Sekahan, and engagement figures are as displayed on October 2, 2026 (JST).
The first impressions from people who tried it were about how it changes the way you talk to an AI. The CTO of an AI business community described keeping a voice line open to his dot.
OpenAIใๆฐใใๅบใใ dot ใฏใใธใงใคใใคใ
— ใฏใใฃใก @ AI Business Lab (@HayattiQ) September 30, 2026
ไฝใใคใใใฃใฆใcall ๆฉ่ฝใใใใ็งๆธใฎใใใซใคใชใใฃใฑใง้ณๅฃฐใงๆ็คบใๅบใใใจใๅบๆฅใใ
ใใใใ็ตใใฃใๆฎต้ใงๅฃฐใงๅ็ญใใฆใใใใฎใงใใใๅ จใฆๅฃ้ ญใงAIใธใฎๆ็คบใ็ตใใฃใฆใใพใใ
ใใใใๆฏๆฅๆฑบใใใใๅฎ็นๅฆ็ใๅ จ้จใใฃใฆใใพใใโฆ pic.twitter.com/9FyujnU2qj
OpenAI’s new dot is seriously impressive. It has a call feature, so you can keep a line open and give instructions by voice like talking to a secretary, and it reports back by voice when the work is done. It also handles the fixed daily routines. It feels like every employee getting a genuinely capable assistant. After one hour I could not go back.
Hayatchi, AI Business Lab, September 30, 2026, translated by Sekahan
An independent developer who has built and sold several services pointed to the single continuous conversation.
ใใฃใใdotsไฝฟใๅงใใใใฉใใใชใๆฅฝใใใ
— ๅ ฅๆฑ ๆ ๅพ (@iritec_jp) September 29, 2026
ChatGPTใ ใจในใฌใใใๅใใใใใใญใธใงใฏใใใจใซไผ่ฉฑใใใใๅบๆฌใฏใใกใใใ่ฉฑใใใใซใใๆใใ ใฃใใ
dotsใฏไผ่ฉฑใใฒใจใคใซใคใชใใฃใฆใใฎใงใใชใใจใชใใไบบใจใใฃใจ่ฉฑใใฆใใๆ่ฆใซ่ฟใใโฆ
I have started using dots and it is a lot of fun. With ChatGPT you split threads or open a conversation per project, and you are the one who goes to talk to it. With dots the conversation is one continuous line, so it feels closer to talking with a person all the time. And it carries that flow straight into Codex and gets the development work done. Talk it through, set the direction, and it does the work. This might be really good.
Shingo Irie, September 30, 2026, translated by Sekahan
A physician and neuroscientist who builds AI tools framed it as a replacement for setups he had already assembled himself.
OpenAIใฎๆฐๆฉ่ฝใdotsใใใใกใใใกใ่ฏใใงใใญ๏ผ
— Daichi Konno / ็ดบ้ ๅคงๅฐ (@_daichikonno) October 1, 2026
ใใใพใงใใผใฝใใซ็งๆธ็ใชๅฝนๅฒใCodexใฎใในใฑใธใฅใผใซใๆฉ่ฝใOpenClawใงใใฃใฆใใใฎใงใใใๅคงๅใdotsใงๅฎ่กๅฏ่ฝใงใ๏ผ
ใใฎใใผใฝใใซ็งๆธใๅฐๅ ฅใใฆใใ้ๅใฎ่ฒ ๆ ใๅๅไปฅไธใซๆธใฃใฆใใใฎใงใใใใใใๆนใใพใจใใใใใฐใจๆใใพใโบ๏ธ pic.twitter.com/LpenOooRbR
OpenAI’s new dots is really good. I had been running a personal-assistant setup with Codex’s schedule feature and OpenClaw, and most of it can now be done with dots. Since I brought in this personal assistant, my administrative load has dropped by more than half, so I hope to write up how I do it.
Daichi Konno, October 1, 2026, translated by Sekahan
Not every first reaction was positive, and one developer changed his mind in public. On launch day he questioned the price for the audience it seemed aimed at, and a day later he quoted his own post with a different verdict.
dotsใฏไปใฎไบบใ่จใฃใฆใใใฉใใฉใคใใฆใผใถใผๅใใชๆฉ่ฝใชใฎใซใใฎใๅคๆฎตใใฉใคใใฆใผใถใผๅใใงใฏใชใใฎใงใฉใใชใ๏ผใฆๆใใฏใใ
— ๆพไธธ ๅฝๅพ(keigo matsumaru) (@k_matsumaru) September 30, 2026
discord็ต็ฑใงไผผใใใใชใทในใใ ไฝใฃใฆใใฎใใใฃใฆdotsใใใชใใฆใใใใใชๆใฏใใ
ใใจใใฃใฑOlusใจSonnetใๅผทใใใ
Dots feels like a feature for light users, as others have said, but the price is not for light users, so I am not sure about it. I had built something similar through Discord, so I am not sure I need dots. And Opus and Sonnet are still just too strong.
Keigo Matsumaru, September 30, 2026, translated by Sekahan
dotsใซใคใใฆใใใชใณใใใใพใใใใๆซใไฝฟใฃใฆใฟใใๆฎ้ใซไพฟๅฉใ ใชใใใคโฆ https://t.co/0cISKSWy1j
— ๆพไธธ ๅฝๅพ(keigo matsumaru) (@k_matsumaru) October 1, 2026
I said that about dots, but after using it for a while, it is honestly just convenient.
Keigo Matsumaru, October 1, 2026, quoting his own post above, translated by Sekahan
The sharpest concern came from a generative AI researcher with a large following, and it was about responsibility rather than speed or price. His reading of the terms of use is his own interpretation, and we have not independently reviewed the clause he refers to.
Dotsใฃใฆๅฉ็จ่ฆ็ดใงใฏใDotsใใๅฎขๆงใซไปฃใใฃใฆ่กใ่ก็บใใใณๅ ฑๆใใใณใณใใณใใซใคใใฆ่ฒฌไปปใ่ฒ ใใใฃใฆๆธใใใฆใฆใฆใผใถใ่ฒฌไปป่ฒ ใใใใใใใใงใใฆDotsใ่ฃใงไฝใใฃใฆใใฎใ่ฟฝใใใญใฐใจใใฆใผใถใฏ่ฆใใชใใใใใชใใงใฆใผใถใ่ฒฌไปปใใฃใฆDotsไฝฟใใฎไธๅฏ่ฝใในใใใธใงใAIใ่ฃใงไฝใใฃใฆใใฎใโฆ
— ใใฟใใ@AI็ ็ฉถ (@umiyuki_ai) October 2, 2026
According to the terms of use, the user is responsible for what Dots does and shares on their behalf. Yet users cannot see logs of what Dots is doing in the background. Taking responsibility for Dots under those conditions is just not realistic, because there is no way to know what the AI is doing behind the scenes. […] No need to panic, but we do have to think about how we live with Dots.
Umiyuki, October 2, 2026, excerpt, translated by Sekahan
Japanese tech headlines ran alongside these posts. Forbes JAPAN’s report carried Sam Altman’s description of Dots as “the true AI he had envisioned”, and PC Watch, a long-running Japanese tech news site, headlined its story “practically a virtual human”. Gizmodo Japan’s headline read “I want to hand over my work. But keep my secrets, please.”, and a note post aimed at B2B marketers argued that managers should decide permission boundaries first, the same question this article examines in the Analysis section.
Taken together, the posts we confirmed show two threads: enthusiasm for an assistant that keeps working and talks back, and a question about who is answerable for what it does when the user cannot see the work. The posts that spread widely were on the enthusiastic side, while the critical posts we found on speed and on Dots adding little over Codex came from smaller accounts and are not quoted here. None of this tells us how common either view is.
Separate Proactive Research From Approved Action
More proactivity means heavier approvals
So far, early users and tech forums are divided on Dots. Even so, its design, which researches and suggests before being asked while acting only within permissions the user sets, clearly distinguishes it from earlier agents that act only when asked. It is a different line of progress from the raw capability that Gemini 4 Argon put forward the same week.
As proactivity grows, what remains for humans is judging exceptions and approving the work that still needs a check before it runs. Whether approvers stay attentive and whether reports are accurate cannot be measured by intelligence scores. That OpenAI, in the same week, held back a model that fell short on staying within scope and reporting accurately shows that the seller, too, has built that weight into its release standards.
The next time someone chooses a product like this, one option is to look, alongside benchmark rankings, at whether the design lets them decide for themselves how much to delegate in advance and where control comes back to them.
Frequently Asked Questions
What is OpenAI’s Dots?
Dots is an always-on AI agent that OpenAI announced on September 29, 2026 at DevDay 2026. Each dot runs on GPT-6 Astra with its own cloud computer and browser, connects to more than 4,000 apps, and keeps working after the user closes ChatGPT.
Can Dots send emails or make purchases without permission?
Not during proactive research. According to OpenAI’s safety post on Dots, the research a dot starts on its own uses read-only tools and cannot send messages, change content in connected apps, or control a browser or computer. Any follow-up action follows the user’s permission rules, which set per action type whether the dot may act without asking, act only when told, ask before acting, or hand the step back to the user, and passes OpenAI’s separate automatic review before it runs. Where an earlier approval already covers an action and no per-action confirmation is required, the dot can proceed on that approval. Purchases with a saved card need the user’s approval, permanent deletion, installing unrecognized software, and new security-sensitive access are confirmed each time, and password changes and transfers between financial accounts are handed back to the user.
How much does Dots cost, and is it available in Japan?
According to OpenAI’s Help Center, the first dot is included at no extra cost in ChatGPT Pro and Business Premium; Pro starts at $100 a month and Business Premium at $125 per user a month ($100 on annual billing), according to plan details reported at launch. According to the same Help Center article, Dots is rolling out to Pro users in markets outside the European Economic Area, Switzerland and the UK, to Business Premium users in all supported ChatGPT regions, and to Enterprise workspaces as a beta that an admin turns on. Japan falls inside both the Pro and the Business Premium coverage, and because the rollout is staged, the Help Center says it may take a few days before a dot appears on an account.
How does Gemini 4 Argon differ from Dots?
Gemini 4 Argon, announced by Google on September 30, 2026, is a model, not an agent product, and represents progress in the intelligence itself. Dots represents progress in how AI is put to work, looking for tasks before being asked.
How did Japanese social media react to OpenAI’s Dots?
Among the Japanese posts we confirmed and quote above, developers who tried it praised the voice call feature, the single continuous conversation that carries into Codex, and the replacement of self-built assistant setups, while one developer first questioned the price and later called it convenient, and an AI researcher raised the question of user responsibility under the terms of use when the background work is not visible in logs. These are individual opinions and experiences, not a survey of Japanese opinion.
Sekahan on YouTube
We publish video summaries of articles like this one, along with short clips built around Japanese reactions.
Reference Links
- Introducing dots | OpenAI
- How we build safety, security, and privacy into dots | OpenAI
- Getting started with your dot | OpenAI Help Center
- GPT-6 Astra System Card | OpenAI
- Gemini 4 Argon: our next era of frontier intelligence | Google
- OpenAI abandons plan to release upcoming model as safety concerns escalate | CNBC
- Dots are OpenAI’s new personal agents | Engadget
- OpenAI launches dots, always-on ChatGPT agents with their own computers | BetaNews
- Introducing Claude Opus 5.5 | Anthropic
- OpenAI’s new agents put safety promises to the test | Axios
