Ongoing topic · Compiled September 30, 2026

OpenAI DevDay 2026: how AI takes on ongoing work

How do dots, Space, Pages and cloud Codex connect ongoing work? Start with the 3D venue demo and product interfaces, then examine model costs, developer tools and the subscription you already use.

At DevDay 2026, Sam Altman stands beside the dots name and four colorful characters.
Dots on stage at DevDay 2026. Photograph from Simon Willison’s live report.Simon Willison · DevDay photograph · Original source ↗View full image ↗
Read by questionContents11

01The change to follow中文

Can the work continue after the request?

Personal assistants, shared documents, cloud coding and third-party apps can sound like unrelated announcements. A useful question connects OpenAI DevDay on September 29: can AI keep work moving after someone leaves the conversation, with results other people can pick up?

Dots takes on ongoing work; Space and Pages hold shared results; Codex Cloud provides development environments; plugins connect outside tools. Model cost and speed shape how far that work can go. Our reading of the combined announcements is that OpenAI is organizing model capabilities into a product experience for continuing collaboration.

There is a history here. DevDay 2025 introduced in-chat apps and the Apps SDK, bringing developers’ services into ChatGPT. In 2026, interfaces gain dedicated workspaces, shared material can be maintained, and supported tasks can begin with new messages or comments. This topic connects official documentation, live observations and later projects, with the base announcements checked through September 30 and later practice dated in each chapter.

Sources and notes

02An ongoing assistant中文

Dots: from a request to an ongoing responsibility

OpenAI’s dots illustration shows Iggy selecting moments from a transcript, drafting show notes and social posts, then revising the materials after feedback.
Official scenario illustration: turn a transcript into clips, show notes and social copy, then carry revisions across the set. This illustrates the product’s intended use; it is not a BitShovel test.OpenAI · dots product illustration · Original source ↗View full image ↗

OpenAI describes dots as Astra-powered agents with their own cloud computer and connected apps. Its launch account includes an early user whose dot noticed a missing publication invoice, prepared it and sent it with his approval. This vendor-reported example shows a concrete sequence: notice an omission, prepare an artifact, then involve the person in the decision.

For a creator, an ongoing assistant could help maintain interview material and drafts; for a small team, it could follow feedback, designs and launch preparation. The useful test is continuity: will accepted wording and settled requirements survive into the next task? These are our suggested uses of the product, to be tested in practice.

Official guidance limits idle proactive research to read-only tools. Task actions depend on connected access, user rules and required confirmations. Initial access covers Pro and Business Premium in eligible markets, plus an admin-enabled beta for Enterprise, Edu and Healthcare. First setup uses the desktop app or desktop browser; messaging can then continue on mobile.

Sources and notes
  • Introducing dots ↗Official product and early exampleCloud computer, proactive research, permissions and initial access.

03Reported use · Continuing tasks中文

Two projects, one dot: where does the handoff stall?

Different states in the author's handoff account
  1. Save, then resume

    On October 2, Work tasks resume from saved state; the previous night's interruption is traced to PC sleep.

  2. Return delegated work

    Two child workers return their checks to the parent Work; full files are not public.

  3. A local wait becomes a full stop

    At 15:58 the dot reports the PC still connected; the account ends with a plan, without an observed restart.

Editorial reading of the author's October 5 follow-up about October 2; not an official rule or a BitShovel rerun. Read the source ↗

On October 1, Creer assigned two local development projects—sluice-gate design material and AI research—to one dot. Existing Markdown, JSON and handoff files helped recover progress before separate Work tasks changed files and ran tests. This is the company's reported use, without full artifacts or a validated engineering result.

In the follow-up, tasks sometimes resumed without another “continue,” and one could advance while the other awaited a decision. Work-path errors and permission handoffs also interrupted progress. Resuming after a notification was opened did not establish that opening it caused the restart; unattended overnight continuation remained unverified.

The useful connection is between saved project state, the next task and decisions returned to the person. Local work also needs its connected computer online and the app open, according to OpenAI. A follow-up being sent and its result reaching the original project are separate observations worth checking.

A third account, published October 5, follows October 2 from morning to 15:58. The local computer had slept overnight. The author later saved work and restart points, updated the app and resumed from them. This documents recovery, not uninterrupted work throughout that period.

The account also reports a parent Work dividing two checks, merging the returned findings into one record and carrying a human answer into later work. That shows a reported handoff beyond sending a continuation request; the full artifacts are not public and reported checks do not validate engineering conclusions.

At 15:58 both projects stopped again. Asked why, the dot said the computer remained connected but partial holds had stopped the projects despite other available work. It proposed resuming; the author had not confirmed an actual restart by the end of the account.

Sources and notes

04A place for shared results中文

Space and Pages: keep work in a shared, editable result

Official ChatGPT Space illustration showing an onboarding prototype on a shared page with two collaborators.
Official Pages collaboration illustration: a prototype, text and collaborators share one page. It shows the editing interface, not a result independently validated by a team.OpenAI · ChatGPT Space product illustration · Original source ↗View full image ↗

Space holds shared pages and files. A Page is a document people can keep writing, visualizing, analyzing and revising. Official examples show teammates editing and commenting in one place, then asking ChatGPT or a dot to continue. The plan can outlast a conversation instead of being scattered across separate chat histories.

For a small event, a useful starting page could contain the venue, budget, program and open tasks. Members revise agreed details, then ask AI to organize next steps. This is our suggested starting scenario. Recurring upkeep also needs instructions about sources and timing; an editable page does not automatically keep every connected source synchronized.

As checked on September 30, Space is available on Pro, Business and Enterprise. Web and desktop support creation and collaboration; mobile initially supports finding, reading and sharing. Collaborative slides and spreadsheets, along with mobile editing, are still forthcoming.

Sources and notes

05October 20 addition · A result people can revise together中文

Slides: what still needs review after shared edits?

From approved material to a deck someone can take over
  1. Agree on material

    Define the audience, sources and open decisions; keep original links.

  2. Revise the shared deck

    Grant appropriate access; identify the element to change and data to retain.

  3. Open the export

    Review figures, charts, text and layout before delivery.

Our suggested starting method based on public guidance, not a UI screenshot or an observed team trial. Read the source ↗

This topic recorded collaborative slides as forthcoming on September 30. The Space documentation read October 20 now describes creating, jointly editing, commenting on and exporting Slides, including direct element edits and requests to ChatGPT. This extends DevDay’s shared workspace. We inspected documentation, without creating, sharing or exporting a deck in an account.

Check the destination first. Space Slides supports same-deck edits and comments that mention ChatGPT or a dot. The shared deck in Space is a different deliverable from a separately generated or opened PowerPoint or Google Slides file. The guide distinguishes mobile creation/viewing from direct element editing on web or desktop, and explains the Windows cloud/web fallback.

Slides currently offers view and edit access, without a separate comment-only role. Linked sources keep their own permissions; material copied or summarized onto a page is visible to that page’s collaborators. Someone may open the deck yet lack access to a source. Confirm Space is enabled and sharing is permitted, then check the recipient’s actual access.

For a small team, we suggest agreeing on figures and open decisions before generating a deck, then making focused edits while preserving approved data. This is an editorial starting method. Web exports include PowerPoint, PDF and Google Slides; desktop opening depends on installed apps. Inspect the exported slides’ figures, line breaks and layout before delivery.

What to check when material enters different results
DestinationHow work continuesCheck before delivery
Slides in SpaceShared edits, comments and element requestsView/edit access and actual content
Linked source materialOriginal permissions remainCan collaborators open the evidence?
Exported filesInspect the delivered formatFigures, charts, wrapping and layout
Sources and notes

06Company use · From records to an application中文

What does AI take on when a family social club expands?

OpenAI's October 1 case study follows Denver family social club The Den as it prepares a second location. Gmail, Slack and Google Drive plugins gather existing records. When a liquor-license hearing moves forward, ChatGPT Work helps find files, identify gaps and convert formats. The team reviews and completes the materials.

The founder reports three hours instead of an expected four days for licensing materials, and two hours instead of three days for grants. These are customer claims in a vendor case, without staff-hour accounting, observation period or full costs. They do not establish profit or approval. A member app and connected sales systems remain plans; the case does not identify dots or Space use.

This offers a concrete reference for ongoing collaboration: retain earlier evidence, organize it around a new requirement and return it for human judgment. It connects to shared pages without validating every newly announced feature. The next question is whether changed records can be carried forward with a visible review history.

Sources and notes
  • OpenAI: The Den's expansion workflow ↗Official customer case · Published October 1Records, plugins and team review. Timings are customer claims; full costs and application outcomes are not disclosed.

07Capability, cost and waiting中文

What do Sol 6.1 and Ultrafast change?

Sol 6.1 changes the capability–cost balance. OpenAI positions it near Astra for coding, computer use and professional work at one fifth of Astra’s standard API input and output rates. It is available in ChatGPT Work, Codex and the API, but not yet ordinary Chat at launch. Our Astra topic compares independent task costs and times to examine which work might move.

Ultrafast changes the speed tier. Official documentation defines Astra’s “up to 8×” in Codex as token generation speed. Searching, tool use, testing and rework still take time. Seeing a first result sooner may encourage more iterations; total time saved needs a complete-task comparison.

At this edition’s check, Astra Ultrafast is available on Pro 500 and eligible Enterprise and Edu plans; Sol 6.1 Ultrafast remains forthcoming. Astra Ultrafast consumes included usage at 8× Standard, while purchased credits and enterprise pay-as-you-go use a 6× billing rate. These are usage multipliers, not speed measurements; API billing is separate.

Sources and notes
  • GPT-6.1 Sol launch ↗Official positioning and accessNear-Astra is vendor positioning; access and unit rates follow the launch.
  • Speed-tier documentation ↗Official docs · checked September 30Generation speed, usage consumption and access are separate.

08How developers connect it中文

Cloud execution and plugins connect the work

Codex Cloud puts code, setup and execution in reusable cloud environments, supporting continuation across devices and shared team setup. This differs from remotely connecting to your own computer. Its environment documentation still lists capability gaps: current Codex cloud environments do not support browser or computer use. Computer use announced for the Agents API should not be assumed available in every cloud task.

Plugin extensions give developers sidebar entries, interactive panels and file viewers within ChatGPT. Design, research and specialized work can stay in an appropriate interface alongside the conversation. Availability varies by client and plan; the documentation still marks web extensions for Free and Go as forthcoming.

MCP Events lets supporting plugins report changes such as new messages or comments, then trigger work according to user instructions. A new task-board item could lead to gathered context and a draft plan. The plugin needs event support and configured actions; connecting an existing plugin alone does not give it continuous monitoring.

Sources and notes

09What was seen on stage中文

A 3D venue, a launch radar—and live interruptions

A DevDay stage screen shows a 3D Fort Mason venue made in Codex, with the task list at left and a web scene containing a stage and chairs at right.
Simon Willison’s photograph of the stage demo shows the venue as a browsable 3D website. It illustrates a creative workflow; public artifacts and full process records are still needed to judge everyday reproducibility.Simon Willison · DevDay photograph · Original source ↗View full image ↗

Simon Willison’s live report describes Romain Huet adding the livestream to a screen inside a 3D venue, and a Sites session turning a launch spreadsheet into a Launch Radar website. One makes rapid iteration visible; the other turns shared material into a browsable tool. These are conference demonstrations, not evidence of long-term reliability.

The same report preserves friction: Dottie paused before responding, and the 3D demo’s voice interaction failed, so the presenter typed instead. Before the keynote, Simon encountered trouble using Codex Cloud for a photo-upload tool and switched tools. That single experience preceded the new demonstration and cannot establish a verdict on the new version.

MicroDuck also appeared on stage; its existing topic explores the hardware and development process. These observations connect two questions: what can AI now help make, and where does a person still take over? The most useful follow-ups will expose inputs, revisions, failures and final artifacts, showing how demonstrated capabilities carry into everyday projects.

Sources and notes

10How subscriptions connect中文

A ChatGPT subscription can travel into other apps

Sign in with ChatGPT separates account sign-in from permission to use your plan allowance. Plus and Pro users can opt into plan usage in eligible participating apps without sharing an API key. Those requests draw from the existing Work/Codex allowance rather than creating a second free pool.

That can change the question “I already subscribe—must I change tools?” Check whether a familiar app participates, then set its weekly cap. The app may still charge for its service, infrastructure or premium features. Developers gain a route to existing users; users may keep familiar tools, while still assessing the total cost.

Pro 500 is the new $500-per-month plan that includes Ultrafast. The help page also confirms reopened Pro 200 sign-ups with a lower allowance than before; eligible existing users retain their previous allowance only through October 29. A personal choice should start with actual work and needed features, rather than treating the fastest announced tier as the default.

Sources and notes

11Bring it back to your own work中文

Start with one recurring piece of work

If you organize events, courses or content, start with real material and see whether Pages makes revisions easier to continue. If you already code with AI, compare local and cloud work, or speed tiers, on the same task. If you build tools, begin with one useful interface or event connection. These are suggested observation paths, not completed user tests.

For users of Opus, Muse or other tools, DevDay provides a reference point: does your task need a different model, a shared editable artifact, or an environment that continues while you are away? These needs do not require replacing everything at once. Identify the part that improves your experience, then consider a switch or combination.

This topic will follow three questions: whether dots reduces repeated explanations over time, how Pages handles conflicting source changes, and what developers actually build with the new interfaces. Reproducible processes and artifacts will return to the relevant sections; if everyday rework cancels out the demonstrated benefits, that will change our assessment too.

What we are following

12What comes next中文

What we are following

  • Does continued use of dots reduce repeated explanations and omissions?
  • How do Pages reconcile shared edits with changes in external sources?
  • Which developer projects expose inputs, rework and final artifacts?
  • When team records change, how can a later application retain a review history?