OpenAI’s dots keep working after you leave

A dot has its own computer. You keep the decisions.

Dots are persistent agents powered by GPT-6 Astra. They have a cloud computer, work through connected apps and carry context between conversations. OpenAI’s launch examples include investigating customer feedback, preparing fixes and turning interview transcripts into clip ideas, show notes and draft posts. These are the company’s proposed workflows, rather than results I have tested.

OpenAI also describes an early tester whose dot noticed a missing invoice, prepared it and sent it after approval. It is a modest example with a clear result: an unfinished piece of work becomes something the person can review. That makes the idea of a persistent agent more concrete than the mascot alone.

Sources: OpenAI [3] · OpenAI Help Center [4]

IMAGE / PRODUCT

Alfred has its own computer. You can take over.

OpenAI’s official demo interface. The Take over control lets the user step into the dot’s computer.

“Always on” has an important boundary

The background research mode can read permitted connected sources and keep notes. Its research tools cannot send messages, change app content or operate a computer. Actions follow separate rules, permissions and review checks. Custom Rules let you define additional boundaries; Activity View shows ongoing work. Connecting your own computer is optional and starts disabled.

That distinction makes the product easier to judge. An agent noticing an unfinished task and an agent acting on your account are different promises. Before delegating an ongoing project, decide which discoveries should become drafts, which changes it may make, and which decisions must come back to you.

Sources: OpenAI Help Center [5] · OpenAI Help Center [4]

GPT-6.1 Sol: the practical upgrade is the economics

GPT-6.1 Sol improves on GPT-6 Sol for coding, computer use and professional work. OpenAI reports performance close to Astra on several evaluations, at lower cost. The launch is in ChatGPT Work, Codex and the API; it is not yet in Chat mode. Calling this a universal “ChatGPT 6.1” upgrade would obscure that difference.

Standard API rates are $2 per million input tokens, $0.10 for cached input and $10 for output. Sol accepts text and images and produces text; image generation uses a separate tool. Its context window is 1.05 million tokens. Above 272K input tokens, input and cache rates double and output costs 1.5 times the base rate.

The attractive possibility is more room for revisions within the same budget. To judge that in practice, use the same repository, brief and requested change for both models. Record the time to a reviewed result, the corrections and the total cost. A cheap draft that needs extensive repair can be an expensive result.

Sources: OpenAI [6] · OpenAI Developers [7]

IMAGE / PRODUCT

GPT-6.1 Sol, introduced at DevDay

Official GPT-6.1 Sol launch graphic from OpenAI’s DevDay recap.

Space gives the work somewhere to live

ChatGPT Space brings files and pages into a shared workspace. People can edit pages together, comment and bring ChatGPT or a dot into the work. Pages can contain research, charts and visualizations. Collaborative slides and spreadsheets are listed as coming soon; mobile currently supports finding, reading and sharing rather than the full editing experience.

For a creative team, I would start with one production brief: references, agreed decisions, deliverables and unresolved questions. The interesting possibility is keeping those pieces together as the project changes. A new comment should update the relevant decision or draft, rather than become another detail someone must remember to copy between tools.

Sources: OpenAI [8]

IMAGE / PRODUCT

Pages, ideas and a shared place to work

Official DevDay product visual. It includes presentation examples; collaborative slides and spreadsheets are still listed as coming soon.

IMAGE / PRODUCT

One page, several collaborators

OpenAI’s example puts teammates and ChatGPT in the same launch plan.

Watch the demo—and the moment it slows down

The live walkthrough starts around 9:40, using a fictional music app called Blossom. The presenter delegates a build, speaks with her dot and brings it into a shared page. Around 12:00, a voice request takes time; later she acknowledges an error in a Codex thread. The presentation shows the direction of the product and some friction alongside it.

That makes this segment worth watching beyond the announcement film. Pay attention to the handoffs: what the presenter delegates, what comes back, and where she checks progress. An agent’s value is easier to understand in that sequence than in a promise that it can handle everything.

Sources: OpenAI / YouTube [2]

Codex can continue in the cloud. Review remains the finish line.

Codex Cloud uses reusable development environments, with a separate workspace for each task. Work can continue while your computer sleeps. The refreshed CLI adds voice steering and an /agents view. Code Review brings pull-request context and diffs into the desktop app; GitHub support is generally available, while GitLab support is in preview. Security Cloud can scan connected repositories and monitor commits.

The cloud environment matters when several changes need the same dependencies and project setup. Instead of rebuilding that context for each task, a team can reuse the environment and review the resulting changes. The practical question is whether a task arrives with enough evidence to make review quicker: a clear diff, relevant checks and an explanation of what changed.

Sources: ChatGPT Learn [9] · ChatGPT Learn [10] · ChatGPT Learn [11] · ChatGPT Learn [12]

IMAGE / PRODUCT

Read the diff. Discuss the change.

The official interface places the code diff and the conversation about it side by side.

The platform changes make the pieces connect

Plugin extensions add native surfaces such as sidebars, panels and file viewers. MCP Events let connected services notify ChatGPT about changes, following subscriptions the user requests. The Agents API now supports an OpenAI-hosted browser for UI tasks. The API itself launched earlier in September; DevDay added computer use, rather than introducing the entire API for the first time.

That combination could make a precise instruction useful over time: when a client leaves a new comment on this approved document, prepare a revision and show me the diff. The event supplies a trigger; the connected tool supplies the document; the review step supplies a stopping point. This is a workflow to build and validate, not something every plugin automatically supports.

Other announcements include Decisions API, Bedrock Managed Agents, Private Intelligence, shareable profiles and OpenAI Marketplace. The official recap links each to its launch status.

Sources: OpenAI Developers [13] · OpenAI Developers [14] · OpenAI Developers [15] · OpenAI [16] · OpenAI [1]

IMAGE / PRODUCT

Canva inside ChatGPT

OpenAI’s DevDay example shows a plugin bringing its own working interface into ChatGPT.

A meeting can become a shared plan

The Meetings plugin turns a conversation into personalized notes and action items in Space. Your own notes stay alongside the summary, which you can share and use for follow-up work. It launches as a macOS beta for Pro and Business; Enterprise has a limited alpha. Other operating systems are still coming.

Team Tasks add shared recurring work: a team can maintain the instructions for a weekly update or a response to a new project event. The team’s configured connections determine which accounts the task uses. For a creative production, that is a useful place for an agreed brief to become a recurring progress update—with a named team responsible for it.

Sources: OpenAI Help Center [21] · ChatGPT Learn [22]

IMAGE / PRODUCT

A recurring task your team can share

The official example shows a scheduled sales summary with follow-up questions in the same workspace.

What your plan actually gets

Access is rolling out gradually. A feature appearing in the keynote does not mean it is enabled in every account. The table separates current availability from previews and upcoming releases; check the linked product documentation before changing a subscription.

A first dot is included with Pro and Business Premium. Talking to it does not use ChatGPT limits; delegated Work and Codex tasks still do.

Pro 500 costs $500 a month and includes Ultrafast. Pro 100 and Pro 200 do not include it. For Astra, the keynote recap claims up to 8× faster token generation in Codex and up to 6× in the API. Those are token-speed claims, not promises that a complete project finishes six or eight times faster. Sol Ultrafast is coming later.

Partner sign-in is a separate change: Plus and Pro users can let participating tools draw on their ChatGPT allowance and set per-app limits. Partner subscriptions and infrastructure charges may remain separate. “Use my ChatGPT plan” does not mean every connected product becomes free.

Sources: OpenAI [3] · OpenAI Help Center [4] · OpenAI [6] · OpenAI [8] · OpenAI Help Center [18] · OpenAI Developers [17] · OpenAI Help Center [19] · OpenAI Developers [13] · OpenAI [1] · ChatGPT Learn [9] · ChatGPT Learn [10] · ChatGPT Learn [11] · ChatGPT Learn [12] · OpenAI Help Center [21] · ChatGPT Learn [22]

IMAGE / PRODUCT

Use your ChatGPT plan in a partner tool

OpenAI’s example shows Devin connecting to a ChatGPT plan. Plan eligibility and usage allowances still apply.

Sol or Sonnet? Compare the job, then the price.

Claude Sonnet 5.5 arrived on 28 September. Anthropic positions it for defined tasks and quick iteration, with Opus 5.5 stronger at complex, open-ended work. Its standard input and output rates match Sol’s. That makes completed-task cost, revision quality and the tools around the model particularly useful things to compare.

There is no defensible overall winner in these launch posts. Anthropic’s published comparisons include GPT-6 Sol, the predecessor of 6.1; OpenAI’s new post compares several results with Opus 5.5. The releases also cite different benchmark versions. Reading those charts as a controlled Sol 6.1-versus-Sonnet 5.5 contest would create a comparison the sources did not perform.

For design and editorial work, I want to see whether a model follows the reference, keeps changes within scope, checks the facts and delivers something usable. Try both on the same materials and keep the first drafts as well as the final versions. That record tells you more about your workflow than a spectacular example chosen by either vendor.

Sources: Anthropic [20] · OpenAI [6] · OpenAI Developers [7] · Anthropic / Claude Platform Docs [23]