OpenAI's DevDay keynote brought more than 20 announcements into one story: AI that keeps working, documents where people and agents collaborate, faster and cheaper model choices, and developer tools for putting agents inside other products. The names are easy to remember. The harder part is knowing what actually shipped on 29 September 2026, what is still a preview, and which plan or region can use it.
Watch the Full OpenAI DevDay 2026 Keynote
Credit: OpenAI's full keynote, published 29 September 2026, with Sam Altman, Romain Huet, Tejal Patwardhan, Holly Li, and the OpenAI team. This guide uses the supplied transcript and OpenAI's official recap, checked on 30 September 2026. Product availability can change after publication.
Launch Status at a Glance
| Area | Available now | Limited or coming later |
|---|---|---|
| Agents and collaboration | Dots rollout for eligible plans and markets; Space pages on supported plans | Specialist dots in enterprise pilots; collaborative slides in the coming weeks |
| Models and speed | GPT-6.1 Sol; Astra Ultrafast API tier; Pro 500 | Decisions API limited preview; Sol Ultrafast coming soon; Private Inference preview planned |
| Developer platform | Codex cloud, refreshed CLI and review, Security Cloud, Agents API computer use | Check plan access, cloud billing, permissions, and rate limits for each surface |
| Plugins and reach | Plugin extensions, improved creation and discovery, sign-in identity | Directory prominence is not guaranteed; plan-backed usage is limited to participating tools |
| Teams and distribution | Team tasks and integrations on eligible business plans; shareable profiles | Meetings plugin is a macOS beta; Marketplace begins with eligible enterprise interest |
These are launch-day categories, not promises that every user sees the same controls. OpenAI's recap lists plan-level access; the linked product guides provide the finer regional, billing, and permission details.
1. Dots and Space: Work That Persists
At 01:26, Sam Altman introduces Dots: an always-on agent powered by GPT-6 Astra. A dot has its own cloud computer, can use connected apps, remembers feedback, and can keep several tasks moving. The launch demo spans a board deck, wedding planning, and a website update. Those are examples of the intended experience, not evidence that any particular account can delegate them safely today.
OpenAI says a dot's proactive research uses read-only access when you are not actively working with it, while action rules and approvals govern other work. Your local computer remains separate unless you connect it. The rollout begins with eligible Pro and Business Premium users; Enterprise, Edu, and Healthcare access is an admin-enabled beta. Crucially for readers in Portugal, the initial personal Pro rollout excludes the EEA, Switzerland, and the UK. The first dot is included for eligible plans, but deeper work has an allowance. An agent that can act still needs narrow permissions and review for consequential actions.
ChatGPT Space is the shared work surface: pages, folders, and team collaboration where ChatGPT and dots can contribute. Pages and Space are available to eligible Pro, Business, and Enterprise users on web and desktop; mobile can find, read, and share pages but not create or edit them yet. OpenAI says collaborative slides are coming in the following weeks. A shared page can expose information copied from a connected source, so review its contents and sharing before inviting others.
2. Models and Speed: Choose by the Job
The model announcement at 18:54 is GPT-6.1 Sol. OpenAI positions it near Astra on complex coding, computer use, and professional tasks at one fifth of Astra's standard input and output token prices. It is available in the API and across eligible ChatGPT plans. That comparison is a published base rate, not proof that every finished task costs one fifth as much; retries and failure rates still matter.
Astra Ultrafast is a premium speed tier. The recap describes up to 8x faster token generation in Codex and up to 6x through the API. Astra Ultrafast is available in the API, and in ChatGPT Work and Codex on Pro 500 and Enterprise; Sol Ultrafast is coming soon. Speed claims describe token generation under specified conditions, not an entire workflow with tools, network time, and human review.
The Pro 500 subscription increases personal usage and includes Ultrafast. Separately, the Decisions API focuses Luna on finite, predefined answers for classification and routing; it is in limited preview. Private Safety Processing is part of the new Private Intelligence story, while Private Inference is a planned preview, not a generally available launch-day option.
3. Codex and Developer APIs: Give Agents a Workbench
From 27:28, the keynote moves into tools for developers. Codex can run cloud tasks with reusable environments, including work started remotely from a phone. The refreshed CLI adds voice steering and an /agents view; the desktop app adds a clearer code-review experience. Codex Security Cloud can scan authorized GitHub repositories, investigate findings, and prepare fixes for review. Cloud scanning has its own billing and access controls; it should not replace tests or human review.
The Agents API now supports computer use and includes multi-agent controls, tool search, tool calling, and context compaction. The important product question is not whether an agent can click a screen in a demo. It is which actions it may take, how they are observed, and what happens when a tool or page behaves unexpectedly. OpenAI also announced Bedrock Managed Agents powered by OpenAI for AWS-native deployment. These are infrastructure choices, not prerequisites for trying the prompts below.
4. Plugins and Distribution: A Route to Real Work
DevDay reopened the question of how third-party products live inside ChatGPT. Plugin extensions can add sidebar homes, interactive panels, and file viewers. OpenAI also introduced easier creation and submission, better discovery, and support for plugins in Sites on eligible work plans. Plugin recommendations are selective; building a plugin does not guarantee prominent placement or customers. Users still choose tools and approve the access they receive.
MCP Events are a proposed specification for event-triggered plugin automations, so treat them as an evolving interoperability surface. Sign in with ChatGPT separates global identity from eligible plan-backed usage in participating apps. The keynote's “bring your tokens” idea does not mean any app can automatically draw unlimited usage from a visitor's subscription.
5. Teams, Plans, and Availability
For teams, the recap adds shared tasks, @ChatGPT in Slack and Microsoft Teams on eligible Business and Enterprise plans, and a Meetings plugin in macOS beta for eligible Pro and Business users. Shareable profiles help people showcase Sites and plugins; the OpenAI Marketplace begins as an enterprise partner channel with eligibility requirements. These launches serve different buyers and should not be rolled into a single “available to everyone” claim.
The keynote closes by inviting builders into the ecosystem. A smaller team's next step is simpler: identify an existing repeated job, test a model or assistant against that job, and keep permissions, spend, and approval proportional to the value. That is more useful than trying every announcement at once.
What to Test First
- Pick one real job. Write down its trigger, input, expected output, owner, and what success looks like.
- Use current access. A ChatGPT conversation, Codex task, or another assistant can prototype the workflow without Dots or a preview API.
- Keep the first run read-only. Use sample data or a manually supplied document; review output before connecting accounts.
- Compare finished work. Track completion, corrections, elapsed time, and cost. For an agent, also log each approval and failure.
Build Something With This
These three prompts are a small experiment for this article. They translate the keynote into something a reader can make now with any capable assistant. Copy one, replace the example with your own context, and keep the first run bounded.
Business: Design one useful agent workflow
Turn a recurring task into a seven-day pilot with a measurable outcome and a clear human approval point.
Personal: Clear one life-admin backlog
Make a repeatable weekly brief without granting an assistant broad account access.
Fun: Build a tiny mission-control game
A playful, single-file project inspired by the keynote demos, with no external services.
Copying a prompt is only a signal of interest, not evidence that someone finished the project. This pilot will inform whether the format deserves a place in future articles.
Keynote Chapters
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:22 | Welcome and introductions | 01:26 | Introducing Dots |
| 08:01 | ChatGPT Space | 09:26 | Dots demonstration |
| 17:28 | Specialist dots and access | 18:54 | Models and speed |
| 23:00 | Research impact | 27:28 | Developer tools |
| 31:30 | Building with new tools | 44:04 | Monetization and distribution |
| 47:54 | Future outlook |
Sources and Link Map
- Event: full keynote and official recap with all announcements and access notes.
- Agents and collaboration: Dots launch, Dots access and safeguards, and Space.
- Models and privacy: GPT-6.1 Sol, Ultrafast API guide, Pro tiers, and Private Safety Processing.
- Developer tools: Agents API and Codex Security Cloud setup.
- Plugins and plans: plugin extensions, plugin platform, Sign in with ChatGPT plan usage, and Meetings plugin.
For commentary on the same launch, read the Matthew Berman recap. For an early hands-on test of Dots across Work, Codex, voice, and scheduling, see the Futurepedia review. For startup opportunities around Dots, the Decisions API, and plugins, see the business-ideas analysis.