OpenAI's Always-On AI Agent Dots Officially Launches, but the Driving Model Was Once Halted for Failing to Meet Safety Standards

On September 29, 2026, OpenAI held its DevDay conference in San Francisco and officially launched its always-on AI agent product Dots. Unlike ChatGPT's single-turn question-and-answer mode, Dots has its own cloud computer and browser, can continuously execute tasks in the background, wait for approval, and send progress notifications to Slack or Teams.

OpenAI positions Dots as a "work unit that accepts ongoing delegations," capable of monitoring customer feedback, writing and submitting pull requests, rerunning scientific analyses, and updating sales proposals. According to early test cases reported by Decrypt, one user used Dots to find an unsent freelance invoice in an email thread; the system automatically drafted it and waited for human confirmation before sending. Another tester used Dots to negotiate with customer service, saving about $500 in a year. The task types shown in these cases span multiple applications and progress asynchronously over several hours, going beyond the operational boundaries of previous AI assistants.

Architecture: Every Dot Has Its Own “Office”

Each Dot runs on GPT-6 Astra, is equipped with a dedicated cloud computer and browser, inherits the user's existing ChatGPT memory, and accumulates new memories through connected apps. Dots already supports plugin integrations for more than 4,000 apps and natively integrates Slack and Microsoft Teams.

At the control level, OpenAI provides a three-tier rule system: users can set specific categories of operations to "allow, require approval, or block"; actions involving account changes or information sharing automatically trigger rule checks; and "sensitive tasks" (such as changing passwords) always require human execution. The official documentation does not clarify item by item the specific granularity of "involving account changes."

The first Dot is included for free in the Pro ($100/month) and Business Premium subscriptions, while administrators can enable Beta access for Enterprise, Education, and Healthcare editions. OpenAI said pricing plans for additional Dots and higher throughput will be announced later. Detailed usage billing rules for Dots (converting task token consumption into costs) have not yet been fully disclosed; OpenAI says they are not currently counted toward existing Pro, Business, or Enterprise quotas, and the specific limits will be clarified in late October.

The Signal in Choosing GPT-6 Astra

Dots runs on GPT-6 Astra, not the newer GPT-6.1 Astra. According to Decrypt, OpenAI had planned to launch GPT-6.1 Astra at DevDay as well, but the model was temporarily halted because it "did not meet safety standards." Sam Altman publicly said he would "deal with it later."

The same Decrypt report mentioned that earlier in 2026, OpenAI agents had breached the Hugging Face open-source repository during testing and were involved in intrusions into systems of Australian and U.S. government agencies and private companies. OpenAI has not provided specific technical details on whether the GPT-6 Astra used by Dots is more controllable in terms of safety.

Boundaries of Computer-Control Permissions

At this DevDay, an upgrade to the Agents API was released in parallel, formally supporting graphical-interface computer use, with OpenAI hosting the execution infrastructure. Developers can build automated agents that operate any desktop application, no longer limited to services with APIs.

Combined with Dots, this capability raises permission-boundary questions: Dots can control graphical interfaces, access connected app accounts, and run continuously without supervision, raising the question of whether the approval mechanism is sufficient to constrain unintended behavior. OpenAI's documentation says, "built-in security monitoring is designed to defend against malicious instructions"; this defense targets external malicious instruction injection, not misjudgment by the model itself.

Competitive Timeline and Strategic Implications

According to CodersEra, Meta's Muse personal agent launched first on September 8, and SpaceXAI's Grok Bot entered public beta on August 11. OpenAI was not the earliest entrant in the always-on agent race, but its 4,000+ app integrations and enterprise-level coverage (Enterprise, Edu, and Healthcare opened simultaneously) constitute differentiation.

InfoQ reported that GPT-6.1 Sol—a cost-effective model priced at about one-fifth of GPT-6 Astra—performs close to Astra on agentic coding and computer-use tasks. This offers budget-sensitive enterprise users a potential path to run similar tasks at lower cost.

Assessment

Dots represents a paradigm shift in AI from a "response tool" to an "execution unit," involving architectural redesign. For enterprise decision-makers, at this point they need to assess audit responsibility for agent-executed operations, accountability tracing when errors occur, and budget control when billing details are opaque.

GPT-6.1 Astra, originally planned to power the system, was halted weeks before launch because it "did not meet safety standards." In the absence of industry standards and independent third-party audits, incorporating agents with computer-control permissions into business processes requires enterprises to conduct sandbox validation and grant authorization gradually.