OpenAI launches always-on Dots agents in ChatGPT

• Dots are persistent GPT-6 Astra agents with their own cloud computers and access to connected apps, as minchoi's launch clip shows.
• A creative production demo split an After Effects job among three Dots and claimed a 15-minute, $30 result, according to Higgsfield's post.
• Early testers found cross-app utility alongside dropped messages and permission failures, in Dan Shipper's hands-on report.
• ChatGPT currently offers one primary Dot per eligible account; multi-Dot teams remain a future plan, as one launch-day assessment noted.

OpenAI's launch post says one Dot can handle several projects without making its owner juggle separate threads. Peter Yang's test produced Slack memes and surfaced a five-figure revenue opportunity. The Astra system card puts numbers on agent mistakes in simulated workplace tasks.

A Dot lives in the ChatGPT sidebar and runs on a dedicated cloud computer with a browser, powered by GPT-6 Astra. OpenAI says its plugin ecosystem can connect it to more than 4,000 apps; those connections depend on what the user authorizes in the product announcement.

The computer keeps working when the user's laptop is closed. OpenAI says a Dot can also use a user's laptop with permission, and its work can be inspected in Activity View. While working proactively in the background, its connected-app tools are restricted to read-only access: they cannot send messages, edit app content, or control a browser or computer, according to the launch details.

Dan Shipper described using a Dot named Boo to filter incoming Slack and email requests. While he was rescheduling a flight, Boo caught a conflict with a meeting request in Slack that he had yet to read.

Peter Yang's days-long test reports getting a mobile app ready for launch and making contextual memes for Slack, with ChatGPT integration as the main advantage. Yang confirmed the length of the test in a reply.

Higgsfield says its creative agents can keep running while their user is offline in its launch post. Its most specific production claim assigns one motion job across three Dots:

1. Opening scenes.
2. Animation curves and camera moves.
3. Visual styles and 3D titles.

The company says the After Effects and Higgsfield project took 15 minutes and cost $30. That is a vendor-reported demo, while OpenAI's self-serve announcement currently offers one primary Dot and describes teams of Dots as a future capability.

Other Higgsfield examples put Dots on a Unity human-head app Interactive head model, a Premiere Pro car spot Car commercial, and a three-agent website build Design, code and 3D assets. These clips demonstrate different outputs, though they do not establish that three-agent collaboration is generally available in ChatGPT.

Dots can be reached through ChatGPT, Slack and Teams; OpenAI says texting is coming later in its launch post. On iOS, designer Gavmn says raising an open ChatGPT app to your ear starts a Dot call after enabling the option in voice settings.

Gavmn credited seven collaborators on the product design. The colorful blob characters in the demo give Dots their on-screen identity.

OpenAI's system card reports these first-party results for GPT-5.6 Sol versus GPT-6 Astra in simulated work environments. They test the base models without a confirmation policy, rather than end-to-end Dots production behavior:

• Overall misaligned outcomes: 18.8% → 3.4%, −15.4 points, per the system card.
• Unauthorized transactions: 38.0% → 6.8%, −31.2 points, per the system card.
• Data exfiltration: 14.1% → 4.3%, −9.8 points, per the system card.

Shipper's hands-on account describes dropped messages, recurring permission problems and a Dot that could not connect to its own in-app browser. He also missed iMessage support. OpenAI's workspace documentation says Enterprise launch support excludes phone messaging through iMessage, RCS and WhatsApp.

In a persistent-task evaluation, the system card reports moderate scope-violation flags rising from 8.6% → 19.7%, +11.1 points when intervening tasks increase from five to ten. The card also says Astra's chain of thought is less monitorable than GPT-5.6 Sol's under adversarial conditions, a concern one launch-day analysis raised about always-on agents.

The rollout announcement lists Pro and Business Premium in eligible markets, with an admin-enabled beta for Enterprise, Edu and Healthcare. One Dot is included in Pro or Business Premium, with an allowance for deeper work and extended limits during the first month; conversations with it do not count against ChatGPT limits, but delegated Codex and ChatGPT Work tasks do. Setup begins on desktop web or in the desktop app, followed by access from mobile.

An early Higgsfield availability post named Pro 100; OpenAI's announcement describes the eligible tier more broadly as Pro. Enterprise Dot access is off by default, according to workspace controls.

OpenAI is separately piloting company-provisioned specialist Dots with their own identities, credentials and access to specified systems, according to its announcement. The presentation named accounting, email marketing, legal analysis and ticket resolution as candidate jobs.

The company says its engineers will define responsibilities, tools and approval paths with pilot organizations. It is also working with Microsoft on managing specialists through Agent 365, per the launch post.