OpenAI Dots: Always-On GPT-6 Astra Agents Ship at DevDay 2026

Update — September 30, 2026: Side-by-side on price, credentials, and incidents: OpenAI Dots vs Grok Bot vs Meta Muse.

September 30, 2026 — OpenAI did ship an always-on ChatGPT agent at DevDay 2026. The product is Dots, not the leaked lowercase "o". Each Dot runs on GPT-6 Astra, gets its own cloud computer and browser, and can plug into 4,000+ apps. You can inspect that computer any time; you can optionally let it use your laptop. The official write-up is Introducing Dots.

The TestingCatalog "o" leak had the shape right and the name wrong. The Get ready teaser never named a SKU. Dots is the shipped form.

Watch: OpenAI’s official Live-from-the-Lab Dots launch film

Open this on openai.com. explainx.ai cannot iframe openai.com (content-security policy), so the official player is a direct link, not an embed.

TL;DR: the questions people asked after the keynote

What can a Dot actually do without you in the tab?

Official product copy is narrower than "it runs my company." A Dot is a persistent agent identity with:

• A dedicated cloud computer and browser (computer-use in the consumer product, not only in the Agents API recap)
• Plugin connections to 4,000+ apps
• Custom Rules, auto-review, and an Activity View
• Safety monitoring that can pause or stop a run
• Optional permission to use your laptop

That is closer to a hosted agent harness with a face than to a longer ChatGPT thread. The interesting split is research vs action.

Proactive research is the background mode OpenAI is willing to run without you watching. It is read-only on connected apps. Official limits: it cannot send messages, cannot change app content, and cannot control the browser or computer. If your mental model is "it booked the flight and emailed the vendor at 3 a.m.," that is not what proactive research is specified to do.

Acting work — the part that would send, edit, or drive a browser — sits behind rules, review, and Activity View. Password changes always stay with the user. That is the right split if you have read MCP security: read and write are different blast radii.

Saved passwords for supported sites are stored so the model does not see them. The cloud computer stays isolated unless you connect yours. Treat "connect my laptop" as a separate trust decision, not a default.

Named examples in official copy — Iggy, Felipe, Todd, Alfred, Jojo — are illustrations, not SKUs you buy. Do not plan procurement around those names.

Who gets Dots, and what does "included" actually mean?

OpenAI did not publish a public extra-dot price list. Use only the rollout language they did publish:

1. Pro and Business Premium in eligible markets — first wave.
2. Enterprise, Edu, and Healthcare — when an admin enables the beta.
3. First dot included at no extra cost.
4. An allowance for deeper work, with extended limits in the first month.
5. Later: more dots, or the ability to scale speed or monthly work.

Free and Plus are not described as launch surfaces. If you are on those plans, this is a waitlist / upgrade story, not a toggle you missed.

Usage accounting is the part builders will get wrong in Slack:

That means a Dot that talks is cheap on the ChatGPT meter, and a Dot that kicks coding or Work jobs is not. If you wire Dots into an engineering org, meter those tasks the same way you already meter Codex — see the DevDay companion posts on GPT-6.1 Sol pricing and the Ultrafast / Pro $500 tier for the rest of the bill, not for Dots SKU prices (those still are not public).

Messaging surfaces: ChatGPT desktop, web, mobile; Slack and Teams; texting soon. A team that lives in Slack can treat a Dot as a channel participant, not only a ChatGPT tab — same distribution pressure a16z described for OpenAI, now with a persistent worker instead of a chat box.

How is this different from Muse or Grok Bot?

Three always-on products shipped in the same quarter. Name them in prose; compare on explainx.ai coverage, not competitor homepages.

The useful comparison is not "which mascot is cuter." It is where credentials live, what background mode is allowed to write, and who pays when the agent starts a coding job. Dots is explicit that passwords are not in the model and that proactive research cannot send. Muse's write-up is explicit about a Sentinel VM and what Meta can still see. Grok Bot's beta is explicit that the agent logs into your tools. Those are three different consent sentences. Pick the one your security review can actually sign.

If you already run always-on coding loops, Dots is not a replacement for Codex persistent mode or loop engineering. It is a ChatGPT-native worker that can start Codex/Work tasks — those still hit those meters.

What are specialist dots, and should a procurement team care yet?

OpenAI describes specialist dots as agents with their own identity, credentials, and systems of record. Early testing named in official copy: procurement, invoices, email marketing, support, and contracting. Microsoft Agent 365 integration is planned, not described as generally available on day one.

Treat specialist dots as an enterprise design — a named worker with a ledger — not as five products on the pricing page. If you are in those functions, the builder question is: does this Dot get a mailbox and an ERP login, or does it only read? Official proactive-research language says read-only. Specialist workflows that pay invoices or send vendor email will need the action path (rules, review, Activity View), not research mode.

Until OpenAI publishes admin docs for specialist identity, do not assume HR / legal / finance sign-off is done. Plan a pilot with one system of record and a human on every send.

What people are asking after the stream

Did the live demos fail, and does that mean Dots is fake?

No. Live demos fail. The product still shipped with a public launch URL and a rollout list.

The clip above is the on-stage flop people clipped. Reporting around the official stream said audio was cut; a second demo also failed. That is embarrassing television. It is not a retraction. Do not mock the presenter. If you evaluate Dots, evaluate your tenant after the rollout reaches you — inspect the cloud computer, walk Activity View, and try a read-only research task before you grant laptop access.

Will OpenAI train on my Dot's research?

Official training split:

• Business, Enterprise, Edu: not used to improve models by default.
• Personal plans: user control.
• OpenAI says it does not train directly on proactive research or the dot’s notes to itself.

That last clause matters. A persistent agent that writes notes to itself is a memory surface. "We do not train on those notes" is a retention / training claim, not a "nobody at OpenAI can ever see operational logs" claim. Read the live privacy FAQ on openai.com when your workspace enables the beta.

Can I run five Dots like a swarm?

Not as a published SKU. Official future language is more dots and scale speed or monthly work. There is no public "buy a swarm" price. If you need parallel workers today, you still compose them with your own harness — or wait for the extra-dot controls.

Is this the same as GPT-6.1 Sol or Ultrafast?

No. Dots is the always-on agent product. GPT-6.1 Sol is a model + price story. Ultrafast and the Pro $500 conversation are speed and subscription stories. Same keynote, different checkouts: Sol launch, Ultrafast / Pro 500, full recap.

Honest limitations

• No public extra-dot prices. "Included first dot" is not a rate card for a fleet.
• Eligible markets are not a country list in the launch post. If Dots is missing, it may be geo, not a bug.
• Proactive research cannot act. Anyone selling "it runs Slack for you overnight" is overselling the research spec.
• Laptop connect expands the blast radius from an isolated cloud box to your machine. Default should stay cloud-only.
• Codex / Work tasks count. A quiet Dot that starts noisy jobs will still surprise finance.
• Specialist + Agent 365 are early / planned. Do not staff a transformation program on those two sentences.
• On-stage reliability was poor in at least two demo beats. Ship ≠ polished live TV.
• Admin beta for Enterprise means your legal and IT still have a gate. Do not enable it in Healthcare because a keynote slide looked friendly.

For tool-using agents, the durable checklist is still MCP security: least privilege, no secrets in the model, audit every write. Dots' password vault and read-only research are aligned with that; connecting the laptop and starting Codex jobs are where teams will skip a review.

What you should do this week

1. Watch the official film — Live-from-the-Lab — then read Introducing Dots for the permission sentences, not the vibe.
2. If you have Pro or Business Premium in an eligible market, create one Dot. Give it a research-only goal. Open Activity View. Do not connect the laptop on day one.
3. If you admin Enterprise / Edu / Healthcare, treat enablement as a beta flag: data-residency, who can attach plugins, and whether Codex/Work spend is capped.
4. Write Custom Rules the way you write a CLAUDE.md: what it may read, what it must ask, what it must never send.
5. Meter Codex/Work separately from "we have a free chat with the Dot."
6. Compare consent language to Muse and Grok Bot if your users will ask "why not the other one?"
7. Follow @explainx_ai and the DevDay hub as admin docs fill in — this post will age on geo lists and extra-dot pricing.

explainx.ai's read

Dots is the consumer-and-prosumer always-on agent OpenAI's September leaks and teasers were pointing at. The hub recap is the right place for Sol, Ultrafast, and the rest of the 20+ announcements. This post is the permission and quota sheet.

The industry move is already happening: persistent identity + a computer + connectors. OpenAI's version is unusually clear that background research is read-only and that dot chat ≠ ChatGPT quota. Those two sentences are the ones to take into a security review. Everything else — specialist dots, Agent 365, more dots later — is a roadmap, not a contract.

If you build agents for a living, Dots does not retire your harness. It raises the default your users will compare you to: a worker that keeps a computer on when the tab is closed.

Related reading

• OpenAI DevDay 2026 — full builder recap
• ChatGPT "o" leak fact-check — name was wrong, shape was right
• Get ready DevDay teaser
• GPT-6.1 Sol launch, pricing, and benchmarks
• Astra Ultrafast and the Pro $500 DevDay tier
• Meta Muse personal agent and Sentinel VM
• Grok Bot persistent agents early beta
• MCP security guide 2026

Official: Introducing Dots · Live-from-the-Lab launch film · DevDay · Livestream

Product rules, rollout, and training language in this post follow OpenAI's Introducing Dots page as of September 30, 2026. Extra-dot prices, country eligibility, and specialist-dot admin controls were not published as a rate card — verify in ChatGPT and workspace admin before you change production access.