OpenAI
DevDay 2026
Announcements and recap
Dots, GPT-6.1 Sol, Codex Cloud and a shared AI workspace. What launched, who can use it, and what is still on the way.
Opening keynote · Sam Altman
Keynote recap20+ announcements · availability varies by plan
The shift: persistent AI workers
OpenAI announced more than 20 updates and reported roughly 1.2 billion weekly users. The six central themes are Dots, GPT-6.1 Sol, Space and Pages, Codex Cloud, plugins, and ChatGPT identity with subscription usage in external tools.
Our reading: ChatGPT is becoming an operating layer for people, applications and agents that keep working between prompts. The key questions are now context, execution, permissions and distribution—not model capability alone.
Compare launch availabilityThe complete announcement map
Launch snapshot, September 30. Announced does not mean enabled for every account.
28 announcements
| Announcement | Status | Access at launch |
|---|---|---|
| DotsPersistent agents with a cloud computer, browser and connected apps. | Rolling out | Pro / Business Premium in eligible markets; admin-enabled Enterprise / Edu / Healthcare beta |
| Specialist dotsDedicated organizational responsibilities, identities and permissions. | Enterprise pilot | Focused enterprise pilots |
| GPT-6.1 SolNear-Astra capability at lower standard token prices. | Available | API, ChatGPT Work and Codex; not regular Chat at launch |
| GPT-6 Astra UltrafastUp to 8× token generation in Codex; up to 6× in the API. | Available | API; Pro 500 and eligible Enterprise / Edu in Work and Codex |
| GPT-6.1 Sol UltrafastAnnounced, with availability to follow. | Coming soon | Coming soon |
| Private IntelligenceZero Data Retention with Private Safety Processing. | Available | Eligible businesses; contact-based access |
| Private InferenceConfidential computing with verifiable controls. | Coming soon | Autumn preview |
| Codex in the cloudReusable cloud development environments and tasks that continue offline. | Available | Plus / Pro / Business / Healthcare / Edu / Enterprise |
| Codex CLIVoice, /agents, delegation, session resume and improved worktrees. | Available | All plans |
| Code ReviewDiff review and automatic first-pass cloud reviews for PRs and MRs. | Available | All plans |
| Codex Security CloudRepository scans, new-commit checks and prepared fixes. | Available | Pro / Business / Enterprise / Edu |
| Decisions APILuna selects from finite answers for classification and routing. | Preview / beta | Limited preview; broader release planned in days |
| Agents API + Computer UseComputer use, multi-agent coordination, tool search and compaction. | Available | API; Pro 500 / Enterprise in Work and Codex |
| Bedrock Managed AgentsOpenAI agent capabilities running within the AWS environment. | Available | Through AWS |
| Plugin extensionsSidebar entries, interactive panels and file viewers. | Available | All plans |
| Plugin Creator & discoveryCreation tools, clearer submissions and improved recommendations. | Available | All plans |
| Plugins in SitesSupported plugins use each teammate’s own connected data and permissions. | Available | Business / Enterprise / Healthcare / Edu |
| MCP EventsConnected-app events can start automations while users are away. | Available | All plans; proposed MCP Events specification |
| ChatGPT SpaceShared knowledge and project work for people, ChatGPT and dots. | Available | Pro / Business / Enterprise |
| PagesCollaborative documents with research, charts, images and comments. | Available | Pro / Business / Enterprise |
| Collaborative SlidesCo-edit with people and agents; export to PowerPoint or Google Slides. | Coming soon | Pro / Business / Enterprise, in the coming weeks |
| Teams & shared tasksShared artifacts and recurring or event-triggered team tasks. | Available | Business / Enterprise |
| @ChatGPT in Slack / TeamsChannel, thread and DM participation with authorized tools. | Available | Business / Enterprise |
| Meetings pluginPersonalized meeting notes and actions, saved to Space. | Preview / beta | Pro / Business on macOS; Enterprise coming soon |
| Shareable profilesShowcase Sites and plugins; skill sharing stays within a workspace. | Available | Free / Go / Plus / Pro / Business / Enterprise |
| Sign in with ChatGPTBring identity and eligible Work / Codex allowance to third-party tools. | Available | Global identity; Plus / Pro usage in participating tools |
| Pro 50025× the Plus usage allowance, including Ultrafast access. | Available | $500/month |
| OpenAI MarketplaceUse part of an OpenAI commitment toward approved partner software. | Preview / beta | Eligible enterprise customers |
Latest updates
Newest first. Each record includes its evidence and status.
7 updates
Decisions API enters preview; Agents API adds computer use
Decisions API uses Luna to choose from predefined answers. Agents API adds computer use alongside orchestration and context tools; Bedrock Managed Agents brings the core capabilities into AWS.
OpenAI · DevDay 2026 RecapPro 500, Marketplace and Private Intelligence
Pro 500 adds a higher usage allowance and Ultrafast. Marketplace lets eligible enterprises spend part of their OpenAI commitment on approved partner software. Private Inference is planned as an autumn preview.
OpenAI · DevDay 2026 RecapCodex expands into cloud development and security
Reusable cloud environments keep tasks running away from your laptop. The refreshed CLI adds voice and an /agents view; Code Review and Codex Security Cloud extend the workflow beyond writing code.
OpenAI · DevDay 2026 RecapDots brings persistent agents to ChatGPT
GPT-6 Astra-powered dots have a cloud computer, browser and persistent context. Rollout starts with eligible Pro and Business Premium users; enterprise beta access requires an administrator.
OpenAI · Introducing dotsOpenAI · DevDay 2026 RecapPlugins gain UI, events and a broader distribution path
Plugin extensions add sidebar panels and file viewers. MCP Events can trigger automations from connected apps. Sign in with ChatGPT separates identity access from eligible subscription usage in participating tools.
OpenAI · DevDay 2026 RecapGPT-6.1 Sol launches; Astra gets Ultrafast
Sol 6.1 is available in the API, Work and Codex, with standard token prices of $2 input, $0.10 cached input and $10 output per million. Sol 6.1 Ultrafast remains a coming-soon announcement.
OpenAI · GPT-6.1 SolOpenAI · Model documentationOpenAI · Ultrafast documentationSpace and Pages make shared work a first-class experience
The collaboration lineup includes team tasks, @ChatGPT in Slack and Teams, and a macOS Meetings beta. Collaborative Slides is scheduled for the coming weeks, rather than available everywhere today.
OpenAI · DevDay 2026 Recap
01 / PERSISTENT AGENTS
Dots: an agent that stays with the work
A dot uses GPT-6 Astra and has its own cloud computer, browser, identity and persistent context. It can keep working while your computer is off and connect to more than 4,000 apps through the plugin ecosystem.
It can communicate through ChatGPT, Slack, Microsoft Teams and voice. A developer workflow might start with recurring user feedback, produce an implemented and tested fix, then bring a PR and a demonstration back for human review.
The official launch name is dots. The earlier o and Aeon rumors are historical context, not separate confirmed products. Specialist dots extend the concept to dedicated enterprise responsibilities and begin with focused pilots.
- 01Event or goal
- 02Gather context
- 03Do the work
- 04Review when needed
- 05Act and follow up
Proactive work has permission boundaries
Proactive research uses restricted read-only app tools by default. Custom Rules can allow an action, require approval or block it. Auto-review examines sensitive actions, and tasks such as changing a password remain with the user. Always-on describes continuity, not unrestricted authority.
02 / INTELLIGENCE × COST × SPEED
GPT-6.1 Sol: the practical model update
Available in API, Work and Codex. Not regular Chat at launch.
What the API gives you
Model ID: gpt-6.1-sol. Maximum output: 128,000 tokens. Reasoning efforts: low, medium, high, xhigh and max. Tool calling uses the Responses API; Chat Completions is supported without tools.
Repeated repository context, system instructions and agent memory can benefit from the lower cache-read rate. Compare the total cost of a completed task, including retries and tools, rather than input prices alone.
Read the pricing boundary
The prices above are Standard rates for prompts up to 272K input tokens. Above 272K, the entire request uses 2× input and cache rates and 1.5× output rates. Cache writes, tools, regional processing and other speed tiers can have separate charges.
What OpenAI’s evaluations show
OpenAI reports near-Astra DeepSWE performance at roughly one-fifth the task cost; on AutomationBench, Sol 6.1 at medium effort leads Opus 5.5 with fallback by 2.2 percentage points at about one-third the cost. On OSWorld 2.0 it is within 2.1 points of Astra at maximum effort, at roughly one-seventh the cost.
Astra remains stronger on the hardest scientific work. These are vendor-reported results under specific evaluation conditions, not guarantees for a production workload.
ULTRAFAST
Latency becomes another choice
Astra Ultrafast is available now: up to 8× faster token generation in Codex, around 300 tokens per second, and up to 6× in the API. Sol 6.1 Ultrafast is coming soon.
service_tier: "ultrafast"For frequent tool calls, OpenAI recommends WebSockets to reduce connection overhead. Faster generation is not the same as an 8× faster completed task.
The platform around the models
Coding, apps, collaboration and distribution are parts of the same shift.
An API for choosing the next action
Give Luna context and a finite set of allowed answers: classify an incoming request, route it to a team, or select the next agent action. The result can feed application logic directly rather than requiring a long explanation to be parsed.
For example, a support system could choose between requesting more information, escalating to a person or preparing a refund for review. Choosing an action does not itself authorize executing it. Access is a limited preview, with broader release planned in the following days.
A development environment that keeps working
Codex Cloud uses reusable environments containing repositories, dependencies, tools, permissions and configuration. Tasks run on managed cloud computers, so closing a laptop does not stop them; users can return from desktop, web or mobile.
The refreshed CLI adds voice, /agents, delegation, session resume and worktree improvements. Code Review connects diffs and discussion with GitHub PRs or GitLab MRs; automatic reviews can start in the cloud. Codex Security Cloud scans repositories and new commits, investigates and deduplicates findings, and prepares fixes.
Apps gain an interface—and a way to start work
Plugin extensions add a sidebar home, interactive panels and file viewers. Plugin Creator, a revised submission flow and better discovery lower the friction between building a plugin and reaching users. Sites can host supported plugins with each member’s own permissions.
MCP Events reverses the usual starting point: a task added to a project board can trigger ChatGPT to read its linked documents and draft a plan while the user is away. OpenAI supports a proposed specification, so this should not be described as a finalized universal standard.
Shared context becomes a place to work
ChatGPT Space brings teammates, ChatGPT and dots around shared project knowledge. Pages combines writing, research, charts, images and comments in a document that both people and agents can revise. This makes the result of a conversation easier to maintain and reuse.
Collaborative Slides is planned for the coming weeks, with concurrent editing and PowerPoint or Google Slides export. Business and Enterprise teams can also share artifacts and instructions for recurring or event-triggered tasks.
Meet people inside their existing workflow
Business and Enterprise users can mention @ChatGPT in Slack or Microsoft Teams channels, threads and DMs. It uses administrator-connected or personally authorized tools; other participants in the conversation do not each need a separate ChatGPT license.
The Meetings plugin starts as a macOS desktop beta for Pro and Business. It saves personalized summaries and action items to Space. OpenAI says the audio is deleted once notes are ready, so the recording cannot then be accessed or replayed. Enterprise support is still upcoming.
Identity and compute allowance travel together
Identity login is available globally, while subscription usage applies separately to eligible Plus and Pro accounts in participating tools. Users can set how much each app consumes. Work and Codex usage still counts toward their plan, rather than becoming an unlimited API balance.
Partners named in the launch include Devin, Notion, Vercel, T3, OpenClaw and Dactyl. Our reading: this can reduce onboarding friction for AI tools and change which part of the experience a third-party developer needs to charge for.
Speed, usage and procurement become products
Pro 500 costs $500 per month and adds roughly 25 times the Plus usage allowance plus Ultrafast. At launch, other self-serve plans cannot unlock Ultrafast just by buying credits. Faster generation also consumes allowance at a different rate; it is not a promise of 25 times as many finished tasks.
Marketplace lets eligible enterprises apply part of an existing OpenAI commitment to approved software. The initial 32 partners span creative tools, customer operations, legal and security, including Figma, Adobe, Salesforce, Harvey and CrowdStrike. It is a procurement channel, distinct from the ChatGPT plugin directory.
More capable agents need clearer boundaries
Zero Data Retention with Private Safety Processing enables automated safety reviews without OpenAI personnel accessing the underlying content. Private Inference, planned as an autumn preview, adds confidential computing and verifiable controls. These are separate capabilities with different availability.
For a production agent, the practical design questions remain concrete: which data can it read, which actions can it take, when does it ask for approval, and who can stop it? A persistent runtime does not remove the need to answer those questions.
Agents API: from orchestration to computer use
Computer use joins multi-agent coordination, tool search, tool calling and context compaction. Explore the foundations below.
Keep the same session
Configuration, conversation and saved work stay with a session. New input can resume an idle session or steer work in progress.
A durable session does not mean a process runs forever.
Seven things worth building with
Our editorial priorities, based on practical product opportunities.
- 01GPT-6.1 Sol
Evaluate quality and total task cost on your own production workload.
- 02Decisions API
Try finite-answer routing before asking for long-form generation.
- 03Agents API + Computer Use
Decide which orchestration and execution infrastructure to keep or delegate.
- 04MCP Events
Map real app events to bounded, reviewable workflows.
- 05Plugin extensions
Design a useful interface alongside conversation.
- 06Sign in with ChatGPT
Separate identity, permission and usage-budget design.
- 07Dots
Study long-term context, triggers and human approval as one product.
What to watch after DevDay
The announced roadmap still contains previews and staged rollouts.
- Decisions API: broader access after limited preview.
- GPT-6.1 Sol Ultrafast: model-specific rollout and final access details.
- Collaborative Slides: delivery over the coming weeks.
- Dots: additional markets and specialist enterprise pilots.
- Private Inference: autumn preview and eligibility.
- Technical session recordings: follow the publication of breakout content.
DevDay Exchanges
Announced local dates in 2026. Venues and programs are provided to approved attendees.
Before the keynote
Preserved for context. These entries reflect what was known before the event; the recap above supersedes their open questions.
Read the six pre-event records
- Confirmed
Schedule checked: keynote at 10 a.m. Pacific
Sam Altman is listed for the opening keynote. OpenAI links to a livestream reminder on its event page. This is our schedule check, not a new announcement today.
OpenAI · DevDay schedule - Confirmed
Durable sessions, subagents and hosted sandboxes are documented
These Agents API capabilities already appear in OpenAI's documentation. This entry records a documentation review, not a DevDay launch or an API release date.
OpenAI · Agent sessionsOpenAI · Multi-agentOpenAI · Hosted sandboxes - Reported leak
Report finds references to an assistant named “o”
TestingCatalog reports an upgrade-screen reference and configuration entries for a display name and “-o” email suffix. The report does not establish a launch date or working email feature.
TestingCatalog · o report, Sep 26 - Speculation
A DevDay launch for “o” remains unconfirmed
The proximity of the report to the event is a reason to watch, not evidence of a scheduled release.
TestingCatalog · o report, Sep 26OpenAI · DevDay schedule - Teaser
Tibo previews work ahead of DevDay
A public mirror of Tibo's post describes an ambitious sprint and changes to how people work. It names no new product. The mirror is linked so the evidence remains inspectable.
Tibo · Sep 24 post (public mirror) - Confirmed
GPT-6 Sol and Luna were released before DevDay
OpenAI's announcement covers GPT-6 Sol and Luna in ChatGPT Work, Codex and the API. These are existing releases, not predictions for the keynote.
OpenAI · GPT-6 Sol and LunaOpenAI · GPT-6 Astra
Frequently asked questions
When did the DevDay keynote take place?
September 29, 2026 at 10:00 a.m. PDT, at Fort Mason in San Francisco. That was September 30 at 01:00 in Beijing, Taipei and Hong Kong. This page now covers the announcements.
Is Dots the official product name?
Yes. OpenAI announced dots as its persistent agent product. The earlier o and Aeon reports were pre-event speculation; they should not be treated as official launch names or proof of the final implementation.
Who can use Dots?
Dots is rolling out to Pro and Business Premium users in eligible markets. Enterprise, Edu and Healthcare users can try the beta if their workspace administrator enables it. Specialist dots begin with focused enterprise pilots.
Is GPT-6.1 Sol available in regular Chat?
At launch it is available through the API, ChatGPT Work and Codex, but not regular Chat. The API model ID is gpt-6.1-sol.
What does GPT-6.1 Sol cost?
Standard rates per million tokens are $2 input, $0.10 cached input and $10 output for prompts up to 272K input tokens. Above 272K, the full request uses 2× input/cache rates and 1.5× output rates. Tool calls and other service tiers may add costs.
Does 8× faster mean every task finishes 8× sooner?
No. OpenAI's figure describes token generation in Codex, not end-to-end task completion. Tool latency and network overhead still matter. Astra Ultrafast is available; Sol 6.1 Ultrafast is still coming soon.
Does Sign in with ChatGPT give every app unlimited usage?
No. Identity login and plan usage are separate capabilities. Subscription allowance sharing applies to eligible Plus and Pro users in participating tools, with user-controlled limits and usage counted against the plan.
Which announcements are still pending?
The main follow-ups are Decisions API's broader release, Sol 6.1 Ultrafast, Collaborative Slides, Private Inference's autumn preview, and expanded Dots access. Announced availability does not guarantee access for every account.