OpenAI DevDay 2026: Every Announcement and What Builders Should Do Next
OpenAI DevDay 2026: Every Announcement and What Builders Should Do Next
OpenAI used DevDay 2026 to launch a model, but the model was not the shape of the event.
The company assembled something larger: a persistent agent with its own computer, a cheaper frontier model, an expensive low-latency lane, a managed developer harness, event-driven plugins, shared workspaces, enterprise automations, portable identity, and a marketplace that can turn OpenAI contract value into third-party software.
Together these releases form a new operating stack for delegated work: intelligence, execution, persistence, software connections, shared state, identity, and procurement.
This article accounts for all 25 named items in OpenAI's official DevDay recap, then separates what builders can use now from what was shown as beta, preview, draft, or future work.
Research cutoff: September 29, 2026 at 12:30 p.m. Pacific. OpenAI had published its recap during the event, but breakouts were still running and session recordings were still pending. Treat this as a complete map of the official announcements available at that cutoff, not a claim that every later demo or session detail had already been documented.
The useful mental model: six layers, not 25 unrelated features
The announcements look scattered when read in launch-note order. Put them into layers and OpenAI's direction becomes clearer.
Every DevDay 2026 announcement, with honest status labels
The table below uses conservative labels. “Live” means an official source says the capability is available, although account-level rollout can still lag. “Beta,” “preview,” “draft,” and “future” mean exactly that. A polished keynote demo is not the same thing as a generally available product.
| # | Announcement | Status at cutoff | Availability and builder meaning |
|---|---|---|---|
| 1 | Dots | Rollout / beta | Primary dots for Pro and Business Premium in eligible markets; Enterprise, Edu, and Healthcare beta when an admin enables it. Specialist dots remain pilots. |
| 2 | GPT-6.1 Sol | Live | API plus ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu. It was not yet an ordinary Chat model. |
| 3 | Ultrafast | Astra live; Sol soon | GPT-6 Astra available through the API and Work/Codex on Pro 500 and Enterprise. GPT-6.1 Sol Ultrafast was announced as coming soon, with no price published at the cutoff. |
| 4 | Private Intelligence | Part live, part future | Private Safety Processing is live for approved ZDR organizations. Private Inference is a preview planned for fall. |
| 5 | Codex in the cloud | Live | Delegated environments and tasks across Plus, Pro, Business, Healthcare, Education, and Enterprise. |
| 6 | Refreshed Codex CLI | Live | Voice steering, subagent controls, worktrees, resume and prompt editing were highlighted for all plans. |
| 7 | Code Review | GitHub GA; GitLab preview | Cloud reviews, diffs, summaries, and review Q&A. Review chat cannot approve, post, or merge by itself. |
| 8 | Codex Security Cloud | Live | GitHub scanning, investigation, deduplication, and proposed fixes for Pro, Business, Enterprise, and Edu. |
| 9 | Decisions API | Limited preview | Luna-based finite-choice decisions from text or images. Broad release was promised in days, but no public endpoint, rates, or price existed at cutoff. |
| 10 | Agents API with computer use | Public beta | The Agents API already launched September 10; DevDay added hosted-browser computer use and product surfaces in Codex/Work. |
| 11 | Bedrock Managed Agents | Available; AWS-controlled | OpenAI's harness and inference run in Amazon Bedrock. AWS determines access, regions, limits, and price. |
| 12 | Plugin extensions | Rolling out | Panels, sidebars, forms, editors, settings, and richer context. Recap says all plans; docs say Free/Go web support is coming soon. |
| 13 | Plugin creation, submission, discovery | Rolling out | Conversational Plugin Creator, redesigned submission, rankings, recommendations, and one directory across ChatGPT and Codex. |
| 14 | Sites can host plugins | Live | Business, Enterprise, Healthcare, and Edu can place plugin experiences inside Sites. |
| 15 | MCP Events | Available integration; draft spec | Available to all plans; ChatGPT supports webhook delivery and callback verification for the draft MCP Events spec on MCP 2.0 (protocol version 2026-07-28), with persistent subscription storage required. |
| 16 | ChatGPT Space | Live with mobile limits | Pro, Business, and Enterprise on desktop/web. Mobile can find, read, and share; creation and editing are coming. |
| 17 | Pages | Live | Shared human-and-AI documents with references, files, comments, instructions, and interactive visuals. |
| 18 | Collaborative slides | Future | Promised for Pro, Business, and Enterprise in the coming weeks. |
| 19 | Teams and shared tasks | Live | Business and Enterprise teams can run cloud tasks using service accounts and configured apps. |
| 20 | @ChatGPT in Slack and Teams | Live | Business and Enterprise collaboration with admin-defined tools and permissions. |
| 21 | Meetings plugin | Beta / alpha | Gradual macOS beta for Pro and Business; Enterprise limited alpha. Other client platforms are coming. |
| 22 | Shareable profiles | Gradual rollout | Web/desktop profile showcases. The recap includes Enterprise, while Help says Enterprise is coming soon. |
| 23 | Sign in with ChatGPT | Live with partners | Identity is available globally in participating tools; Plus/Pro plan usage is limited to participating integrations, consumes Codex or ChatGPT Work allowance, and requires separate consent. |
| 24 | Pro 500 | Live | $500 per month, 25 times the Plus allowance according to the recap, with Astra Ultrafast included at launch. |
| 25 | OpenAI Marketplace | Interest-led rollout | Eligible enterprises can express interest in applying OpenAI commitment value to approved software from the first 32 partners. |
“Announced” spans a callable API, controlled beta, protocol proposal, and future promise. Flatten those categories and a team can build a roadmap around products it cannot deploy.
The six stories inside the event
1. Dots make persistence a product primitive
OpenAI's Dots launch describes a GPT-6 Astra-powered agent with its own isolated cloud computer and browser. It can continue working, remember context, and use thousands of apps through plugins. The companion safety explanation is as important as the demo: credentials are separated from the model, risky actions trigger approvals or handoffs, proactive research is read-only, and an external Auto-review layer watches the environment.
The key insight is that a dot is not a smarter chat tab. It is a persistent digital principal. Once software can remember, hold access, and act while a person is absent, identity and revocation become core product design.
Read the full analysis: OpenAI Dots: What the Always-On AI Agent Can Do.
2. GPT-6.1 Sol changes the routing boundary
GPT-6.1 Sol offers a 1.05-million-token context window and 128,000-token maximum output. For prompts with up to 272K input tokens, standard API rates are $2 per million input tokens, $0.10 cached input, $2.50 cache writes, and $10 output. Long requests reprice the entire request, not only the excess. OpenAI's launch benchmarks are promising, but they remain vendor-run evidence.
The hidden price story is caching. A repeated token costs five percent of a fresh input token. The hidden risk is the 272K cliff. Good architecture can matter more than a small benchmark lead.
Read: GPT-6.1 Sol: Benchmarks, Pricing, API, and the New Routing Math.
3. Codex and the Agents API expose one harness in several places
Codex cloud provides durable remote environments. The refreshed CLI, Code Review, and Security Cloud cover more of the delivery loop. The Agents API, already public beta before DevDay, exposes the harness to developers; DevDay added hosted-browser computer use. Bedrock Managed Agents place a related stack inside AWS.
This is OpenAI selling the same execution idea at three levels: finished product, programmable service, and cloud-provider deployment.
Read the Codex deep dive: OpenAI Codex Cloud, CLI, Code Review, and Security.
Read the API deep dive: OpenAI Decisions API and Computer-Using Agents.
4. Plugins become applications and event sources
Plugin extensions add panels, editors, forms, file experiences, and settings. MCP Events adds webhook subscriptions, although the specification remains a draft.
That combination changes the plugin from a tool the model occasionally calls into a UI surface that can also initiate work. The permission model must now cover rendering, context exchange, subscriptions, retries, and side effects.
Read: OpenAI Plugins, Sign in with ChatGPT, and the Marketplace.
5. Space and Pages turn chat into shared operating state
Space and Pages move useful work out of private threads and into artifacts that people and agents can edit together. Teams, tasks, Slack and Microsoft Teams integration, Meetings, and future collaborative slides surround those artifacts with execution and communication.
The crucial permission detail: connected sources retain their own access controls, but information copied or summarized onto a shared Page becomes visible to Page viewers. Sharing a Page does not expose private chats or memory. That boundary should be part of every rollout guide.
Read: ChatGPT Space, Pages, and Team Tasks.
6. Identity and procurement become distribution
Sign in with ChatGPT separates identity from plugin permissions and from permission to spend a plan allowance. That is sensible. The Marketplace then adds a commercial path: approved enterprise customers can express interest in applying part of an OpenAI commitment to partner software.
This is not a public consumer app store. It is closer to enterprise distribution through identity, existing budget, and procurement relationships. Pro 500 adds a high-allowance individual tier at $500 per month and includes Astra Ultrafast at launch.
Those commercial layers are covered with the plugin platform in OpenAI Plugins, Sign in with ChatGPT, and the Marketplace.
What I would do next as a builder
Choose a task where retries or human review are costly. Compare accepted outcomes, task cost, latency, and long-context behavior against your present route.
Use a dot, Team Task, or MCP Event in a low-risk environment. Define identity, allowed tools, approvals, stop conditions, and revocation before adding reach.
Track Decisions API, Private Inference, collaborative slides, and specialist dots, but wait for public contracts, prices, limits, and failure behavior.
Choose one boundary that changed. Test Sol if model spend dominates; Space if work loses context; MCP Events if polling blocks a workflow; or the agent runtimes if execution needs a durable home. For persistent agents, begin with authority rather than personality.
Three non-obvious conclusions
The control plane may matter more than the model lead
Model rankings move quickly. Identity, permissions, state, distribution, review, and procurement are stickier. OpenAI wants builders anchored to those layers even when the best subtask model changes.
Raw APIs and finished agents will have different safety contracts
Dots add product controls around credentials, approvals, and review. The Agents API warns that origin approval does not guarantee confirmation for every action. Lower-level primitives transfer more responsibility to the builder.
Shared artifacts are the bridge between people and agents
Pages can carry evidence, instructions, comments, files, and task state across people and agents. The artifact, not the conversation, becomes the natural coordination unit for approvals and provenance.
FAQ
When and where was OpenAI DevDay 2026?
It took place Tuesday, September 29, 2026 at Fort Mason in San Francisco. The official event site listed an opening keynote at 10:00 a.m. Pacific, breakouts through the afternoon, a closing session at 4:00 p.m., and recordings to follow.
Were all 25 announcements generally available on launch day?
No. The set included live products, gradual rollouts, betas, a limited preview, a draft protocol, and future releases. The status table above preserves those distinctions.
Is GPT-6.1 Sol available in normal ChatGPT chat?
At the research cutoff, OpenAI documented it in the API and in ChatGPT Work and Codex for eligible paid and organizational plans, not as a normal Chat model.
Did OpenAI launch the Agents API at DevDay?
No. OpenAI introduced it in public beta on September 10. DevDay added computer use with an OpenAI-hosted browser and expanded how the agent runtime appears across products.
Is MCP Events production-ready?
OpenAI documents it as a draft extension to MCP 2026-07-28. It is important directionally, but production teams should expect protocol and interoperability changes and should build idempotency and recovery around every webhook.
Is the OpenAI Marketplace open to anyone?
Not according to the launch material. OpenAI named an initial set of partners and invited eligible enterprise customers to express interest. Public catalog pricing and broad self-service terms were not available at the cutoff.
Final take
OpenAI DevDay 2026 is best read as a map of where the company wants delegated work to live.
The model reasons. The harness acts. The dot persists. Plugins connect. Pages preserve shared state. Sign-in carries identity. Plans and the Marketplace route spending. Private processing and product-level controls try to make the whole system acceptable to organizations that cannot treat an agent like an enthusiastic browser extension.
Some pieces are ready. Some are carefully gated. Some are still proposals with good demos attached.
Builders should resist both easy reactions: neither “everything changed today” nor “this was just another keynote” is accurate. The practical response is to use the status map, test one real boundary at a time, and keep authority, evidence, cost, and rollback visible.
The model release will get most of the attention. The platform assembled around it is what may last.