OpenAI Launches Dots, Always-On GPT-6 Astra Agents Starting With $100 Pro Access

OpenAI introduced Dots on September 29

OpenAI announced Dots on September 29, 2026 as proactive assistants for projects and everyday work. The product documentation is more specific: a dot is an always-on agent powered by GPT-6 Astra, with its own cloud computer and browser. It can continue working while the user’s device is off and return with results or decisions that need attention.

This is a ChatGPT product rollout, not a new API model release. OpenAI does not document a Dots API identifier, per-token rate or general developer endpoint in the reviewed launch material. Builders should not translate “powered by GPT-6 Astra” into an assumption that a dot’s orchestration, memory, permissions or cloud computer can be reproduced by calling an Astra model endpoint.

Access starts with Pro, Business Premium and Enterprise

OpenAI says Dots are rolling out gradually, so plan eligibility does not guarantee immediate access. Pro 100, Pro 200 and Pro 500 users must be over 18 and outside the European Economic Area, United Kingdom and Switzerland. Business Premium and Enterprise are rolling out worldwide; Enterprise workspaces have Dots off by default until an administrator enables them.

The current pricing page lists the three Pro tiers at $100, $200 and $500 per month. OpenAI does not state a separate Dots fee, a per-task Dots rate or public numeric Business Premium and Enterprise prices in the reviewed pages. Conversations with a dot do not count against ChatGPT usage limits, but tasks it starts or manages in Work or Codex consume those products’ limits as usual. OpenAI also says plans include an allowance for deeper work, with extended limits for the first month, without quantifying that allowance on the Dots page.

A dot combines continuity, delegation and a cloud computer

A dot can keep more than one responsibility moving, decide when to pause and resume work, and divide work among background agents. It can create visible cloud threads, use a configured Codex cloud environment for coding, or create and continue eligible Work or Codex tasks on one connected personal computer. Fixed-time work still requires a saved schedule; event monitoring requires a supported source and an explicit request. Merely connecting Slack does not create a monitor.

Its continuity comes from the active conversation, relevant ChatGPT memory and separate saved notes about preferences, decisions and ongoing work. Those notes are not a full transcript and are not necessarily changed by changing ChatGPT saved-memory settings. The same dot can be reached through ChatGPT, Slack and Microsoft Teams; visible conversations stay separate even though relevant context can carry across them.

Connections grant different kinds of authority

The cloud computer has its own files, software and browser sessions. A user can take over its mouse and keyboard for a login or other step, then return control. Separately, one personal computer can be connected so the dot can work with local files, apps and tasks while that computer is online and the ChatGPT desktop app is open. A cloud session does not inherit logins from the user’s browser, and a local-computer connection is not the same permission as Codex access or Work Sync.

Supported plugins can expose services such as Gmail, Google Drive and GitHub according to the connected account and allowed actions. Messaging, plugins and computer access are independent: adding a dot to Slack does not connect email or a laptop. That separation is useful only if teams preserve it in deployment records instead of granting every connection during setup.

The control model has more than one stop button

OpenAI says automatic review checks actions against instructions, app permissions, custom rules and built-in safety requirements. An action can proceed, require approval or be handed back to the user. Custom rules can allow an action without asking, require an explicit instruction, ask immediately before acting or hand it off. OpenAI also warns that these rules are instructions the dot tries to follow and that it can make mistakes; they do not override built-in safeguards or grant an app connection.

Operationally, Pause stops the current main task only. Delegated tasks must be inspected and stopped separately in Activity, and recurring work must be disabled or deleted in Scheduled. Ending a call also does not necessarily stop assigned work, and stopping work does not undo actions already taken. Teams evaluating Dots should test all three shutdown paths and confirm the external state after each one.

What teams should do next

Eligible individuals can start with one reversible responsibility that uses read-only sources and produces drafts for review. Record the plan, region, connected channels, plugins, computer access, schedules, approval rules and where delegated work runs. Then test stale context, cross-channel disclosure, expired sessions, a blocked cloud browser, an offline local computer, approval requests and each stop path before allowing writes.

Business and Enterprise buyers should wait for tenant access and admin controls to be visible in their own workspace, then verify contract terms, retention, audit events, identity lifecycle, usage limits and incident response. Reject unattended authority for payments, irreversible deletion, credential changes or external publication until deterministic controls outside the model bound the action and a complete rollback path has been exercised. The product is available to some accounts now; its reliability and governance fit remain workload-specific and unverified here.

Disclosures

AccessAllGPT did not receive OpenAI access, credits, a subscription, briefing, demo, review or compensation for this article. OpenAI did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI or organizations cited. Publication-wide relationships are listed on the disclosures page.