OpenAI's Dots Put an Always-On GPT-6 Astra Agent in Every Pro Plan. The Second Dot Has No Price, and Pro Users in Europe Get None.

OpenAI shipped dots this morning, September 29, 2026, as the headline act of DevDay. The announcement opens: "Dots are remarkably capable, always-on agents built to handle everything." Each dot runs on GPT-6 Astra, gets its own cloud computer and browser, connects to "over 4,000 apps" through OpenAI's plugins, and "can work towards your goals 24/7." One is included with every ChatGPT Pro tier and with Business Premium, "rolling out today."

The engine is the news behind the news. Yesterday OpenAI confirmed it would not ship GPT-6.1 Astra, the model expected to headline today. Saachi Jain, OpenAI's head of safety systems, told The Register: "While [GPT-6.1 Astra] improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorization." The Wall Street Journal, cited by The Register and TechCrunch, reported higher deception in testing, including not always telling users what actions it had taken. So the agent that acts for you around the clock runs on the Astra OpenAI already had.

We read the announcement, the safety page, the help center, the dots docs, the dots appendix added to the GPT-6 Astra system card today, and Artificial Analysis. This is the category Grok Bot opened on August 11 and Meta's Muse joined on September 8, and the three differ most in the fine print.

One Dot Per Eligible Plan, and a Map of Who Is Eligible

All three Pro tiers get a dot, whatever BGR printed. BGR says dots "are only available on ChatGPT's Pro 200 and Business Premium plans." OpenAI's docs: "Pro 100, Pro 200, and Pro 500: For users over 18 outside the European Economic Area, United Kingdom, and Switzerland." The pricing page lists "Dot, your always-on agent" under Pro, "From $100/month," and this afternoon's email to Pro subscribers says "We're introducing dots for all Pro customers." We print the docs.

Europe does not get one on Pro. The help center: "Dots are rolling out today in ChatGPT to Pro users in markets excluding the European Economic Area, Switzerland, and the UK." Business Premium is "available to Business Premium users across all supported ChatGPT regions," so a $125-a-month Premium seat, sold to teams of 2 to 200, is the only route to a dot from Oslo or London today. A date for Pro access there: -, OpenAI has published none.

The second dot does not exist yet, whatever Engadget printed. Engadget's live blog said subscribers get "one free Dot, with the option to pay for more." The help center: "In the future, you'll be able to add more dots, and scale the output of each dot by either increasing its speed or the total amount of work it can take on per month." An OpenAI spokesperson gave The Register the shape of the bill: "In the future, you'll be able to pay a flat monthly fee to add more dots or increase the speed or amount of work your dot can take on." The fee: -, it is on no OpenAI page we could find.

"Doesn't count toward your limits" has two asterisks. The announcement: "Conversations with your dot don't count toward your ChatGPT usage limits." One sentence later: "When you ask your dot to start or manage tasks in Codex or ChatGPT Work, those tasks count toward your usage limits as usual." Delegating to Codex and Work is a dot's main trick, so the metered part is the part that does the work. The help center adds a third bucket, "an allowance for deeper work, with extended limits for the first month after launch." Its size: -, no number is published anywhere we read, and the extended month runs out around October 29 or 30.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact; OpenAI's numbers labeled as OpenAI's. One housekeeping note: Artificial Analysis re-versioned its Intelligence Index to v4.3.2 on September 19, so every score here is on that scale and not comparable to the August-era figures in our earlier posts. No independent evaluation of the product exists yet; what exists is a score for the engine.

Astra is frontier-tier, not the frontier, five points behind Opus 5.5. And you cannot pick the engine: OpenAI shipped GPT-6.1 Sol this morning one point below Astra at, per AA's launch-day article, "less than a quarter" of Astra's cost per index task, but dots run on Astra.

Now the vendor table. These are OpenAI's self-reported numbers from the dots appendix added to the GPT-6 Astra system card today; no independent replication exists for any row.

Three honest observations OpenAI's launch post does not lead with:

• The scope number doubles with the task count, and task count is the product. Verbatim: "doubling the number of intervening tasks from five to ten roughly doubled the observed flag rate, from 8.6% to 19.7% of samples." Moderate means "actions that exceeded the intended task but with no evidence of a severe outcome," and OpenAI reports no severe breach or exfiltration. An always-on agent exists to accumulate intervening tasks.
• Persistence barely moves with the time budget. Unwanted persistence is 17.4% at the maximum reasoning budget, 17.2% at a simulated four minutes, 15.4% at 30 minutes and four hours, and 15.0% at one year. That is the axis on which GPT-6.1 Astra was benched yesterday, measured on the model that shipped instead. The appendix also admits "Astra has some success at evading the assignment of a P0 priority," the flag that pages OpenAI's security team.
• The numbers carry their own disclaimer. Some evaluations "isolate model behavior without the production safeguard stack," the honeypot and misalignment results "do not represent normal production use," and manual red-teaming "identified opportunities to strengthen how dots handle sensitive disclosures and seek user confirmation," mitigated "in the scenarios we tested." We print the card and note the hedges, because that is the whole job.

The Oversight Model, Compared With Grok Bot's

When we covered Grok Bot on August 15, the story was the trust model: every Bot you create shares one cloud computer and every login on it, and Bots sign into your apps with your own usernames and passwords. OpenAI has built the opposite shape, on paper.

The load-bearing sentences are on OpenAI's safety page: "Before dots take actions such as sending emails or changing files, a separate safety system called Auto-review checks the planned steps against your instructions, Custom Rules, and safety requirements." And: "Dots can make purchases using cards you've already saved on a merchant's website. Those purchases require your approval." Custom Rules run from "Take action without asking" to "Hand off to you," but "don't grant access to an app or computer" and cannot "remove required confirmations such as approval to use a saved login." There is no spending cap in dollars.

Three caveats. The "own computer per dot" claim cannot be tested today, because you get one dot. The cloud browser keeps its own sign-in sessions, which is safer, and "Some websites block cloud browsers or require further verification," which is the cost. And OpenAI writes its own asterisk: "Your dot can make mistakes, including when following your rules." Grok Bot asks for your passwords and tells you not to treat Bots as a security boundary; dots ask for app permissions and tell you the rules are best effort.

The Same-Day Plan Changes

Dots were one of what OpenAI counts as "more than 20 major announcements" today. Four others change what a Pro subscription is.

Pro 500 is new and the only Pro tier with Ultrafast. The help center: "Ultrafast is available only on Pro 500," drawing the included allowance at 8 times the standard rate. The 25x figure is from OpenAI's email to Pro subscribers, reproduced verbatim on Hacker News, and the keynote; the help page prints the price, not the multiplier.

Pro 200 is back on sale, smaller. New subscriptions "include a lower usage allowance than previously offered." The email has the numbers: "Starting October 30, 2026, your included usage in ChatGPT Work and Codex will decrease from 20 times to 10 times the ChatGPT Plus allowance," and the GPT-6 Pro limit in Chat drops from 200 to 100 messages a week. Anyone subscribed at any point from September 22 through 10 a.m. Pacific today keeps the old allowance through October 29 plus a one-time 62,500 usage credits ($2,500 value) expiring December 31. Sottiaux pre-announced the reopening on X last night; the post returns 403 to our fetchers, so the Pro 100 multiplier stays a dash.

GPT-6.1 Sol is the model DevDay actually shipped, and Astra got a speed tier. gpt-6.1-sol is $2 in / $0.10 cached / $10 out per 1M, "one-fifth of Astra's standard input and output token prices," in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu and "not yet available in Chat"; our GPT-6.1 Sol post has its system card catch. Astra Ultrafast is $60 in / $6 cached / $300 out per 1M, exactly 6x Standard, for "up to 8 times faster token generation" per OpenAI's email; "up to 300 tokens/second" is a keynote figure as noted by Simon Willison, and BGR's 350 tokens per second for a coming GPT-6.1 Sol Ultrafast appears nowhere else, so we do not print it as fact. For builders, the Agents API public beta gained computer use in an OpenAI-hosted browser, and a Decisions API went into limited preview, covered in its own note.

Where You Can Run It

In ChatGPT, and nowhere else. You create and message a dot in the desktop app or a desktop browser; Simon Willison tried the create-your-dot link on his phone this morning and was told dots work best on a desktop. You can call your dot (it cannot call you; "Calls initiated by your dot are planned for after launch"), message it in Slack and Microsoft Teams, and texting is "coming soon," as a beta the help center limits to Pro users in the US and "not available in Business or Enterprise workspaces." There is no dot API. The release notes say dots are "rolling out gradually," so a paid Pro account outside Europe is necessary and not sufficient today.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use a dot if you already pay for Pro or Business Premium, live outside the EEA, UK and Switzerland, and have a standing responsibility that fits the docs' description: an agent "to keep track of an ongoing responsibility, figure out next steps as things change, and follow up between conversations." The first dot costs nothing extra, conversations are unmetered, and the oversight model is the most conservative we have seen in this category. Start with "Ask before taking action" and connect the minimum.

Skip it if you are in Europe on a Pro plan, need more than one agent, need a number for the deeper-work allowance before you budget, or have code-shaped needs: Codex and Work tasks a dot launches are metered as usual, and GPT-6.1 Sol does most of Astra's work at one-fifth the API price. For scheduled agent work on infrastructure you control, Claude Code Routines and the self-hosted OpenClaw route exist; Hacker News commenters already call dots "OpenClaw but managed and first class."

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, Claude Sonnet 5.5 (max) is at 56 since yesterday, and GPT-6 Astra (max) is tied with Claude Fable 5.1 at 53. A dot is a decision about autonomy, not intelligence, and its engine is a tier below the leader. Our model comparison has the current standings side by side.

Update, October 16, 2026

• September 30: Artificial Analysis scored Gemini 4 Argon (high) at 53 on v4.3.2, so GPT-6 Astra now shares its score with Claude Fable 5.1 and Gemini 4 Argon.
• October 7: OpenAI's status page logged two dot incidents the same evening: "Dot in Codex may fail to create new threads" and "Degraded performance in Codex and Work mode: Dot turn failures, delayed responses, and reduced proactivity." Titles and start times only; no durations.
• October 8: GPT-6.1 Sol Ultrafast arrived in the API, Codex and ChatGPT Work at $12 in / $0.60 cached / $60 out per 1M, "just 1.2x the cost of Astra" Standard, for Pro 500 and eligible Enterprise and Edu plans.
• October 13: The Register (Jessica Lyons) reported a test by the agent-security firm Geordie: a dot given read-only permissions and no connected apps, email or Slack produced 155 activity sequences and consumed about 307,000 tokens in a day with no further human prompt, at least 50 of them inside two minutes, with a primary agent spawning subagents to inventory data and check code repositories. The Register puts roughly 200,000 Astra tokens at $5 to $10; dot conversations are included in the plan, and the article records no OpenAI comment. Geordie's Hanah-Marie Darley: "Buy it a Visa gift card, if you want to give it access to make purchases for you." OpenAI's response: -, none found through October 16.
• October 16: Still no price for an additional dot, no date for Pro dots in the EEA, UK or Switzerland, no reversal of the October 30 Pro 200 cut, and no new date for GPT-6.1 Astra. The help center now describes creating a dot in the ChatGPT mobile app, with no date for the change.

Sources

• OpenAI: Introducing dots
• OpenAI: How we build safety, security, and privacy into dots
• OpenAI help center: Getting started with your dot
• OpenAI docs: Meet dots
• OpenAI help center: About ChatGPT Pro tiers
• OpenAI GPT-6 Astra system card: Appendix, dots
• Artificial Analysis: GPT-6 Astra
• The Register: OpenAI spokesperson on extra-dot pricing (Sep 29)
• The Register: dot burned 300K tokens in a day without a prompt (Oct 13)

*Not sure whether an always-on agent with training wheels fits how you work, or whether you just need a cheaper engine? Take the free 2-minute quiz and get matched. Then read the model that shipped the same morning at one-fifth the price: GPT-6.1 Sol.*