OpenAI DevDay 2026: Everything OpenAI Announced, From GPT-6.1 Sol to Dots and Codex Cloud

OpenAI DevDay 2026 was much bigger than a single model launch

OpenAI’s annual developer conference took place on September 29 at Fort Mason in San Francisco, and this year the company used the event to show a much broader vision for what it wants developers and users to build with AI.

OpenAI says DevDay 2026 was its biggest DevDay yet, with more than 20 major announcements spanning ChatGPT, Codex, models, APIs, plugins and new ways for people and AI agents to work together. The company also said ChatGPT now reaches 1.2 billion weekly users.

The event was not centered on one giant frontier-model reveal.

Instead, OpenAI pushed several pieces of the AI stack forward at the same time:

• A new lower-cost GPT-6 model

• Always-on AI agents called Dots

• Cloud-based Codex environments

• Computer use in the Agents API

• A new Decisions API

• AI-powered security tools

• Plugin extensions inside ChatGPT

• Shared workspaces and Pages

• Sign in with ChatGPT

• A new $500 Pro tier

• An enterprise software marketplace

Taken together, the announcements point toward a larger change: OpenAI wants ChatGPT to become not only a place where people ask questions, but a platform where people, applications and persistent AI agents can work together.

1. GPT-6.1 Sol is the headline model launch

The most important developer-facing model announcement was GPT-6.1 Sol.

OpenAI describes it as a major upgrade to GPT-6 Sol that approaches GPT-6 Astra’s performance on agentic coding, computer use and professional work while using substantially less compute and costing much less through the API.

OpenAI says GPT-6.1 Sol improves on complex tasks including:

• Software development and debugging

• Professional document analysis

• Multi-step business automation

• Computer interaction

• Scientific workflows

• Factual accuracy

The standard API price is:

These are standard short-context API prices published by OpenAI.

GPT-6.1 Sol is available through the API and is rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. It is not currently available in the regular ChatGPT Chat interface.

How close is GPT-6.1 Sol to Astra?

OpenAI's own evaluations are unusually focused on agentic work.

On DeepSWE v1.1, a software-engineering benchmark involving long-horizon tasks in real codebases, OpenAI says GPT-6.1 Sol matches GPT-6 Astra at roughly one-fifth the cost on the tested setup.

On OSWorld 2.0, which measures computer-use workflows, OpenAI says Sol comes within 2.1 percentage points of Astra at maximum reasoning effort while costing roughly one-seventh as much per task in the tested comparison.

On AutomationBench, OpenAI reports that Sol scored 2.2 percentage points above Opus 5.5 at medium reasoning effort and 4.8 points higher than GPT-6 Sol in the same setting.

OpenAI also says GPT-6.1 Sol reduced the share of difficult test responses containing a factual error from 11.4% to 7.7% at low reasoning effort.

These are OpenAI-run evaluations, and OpenAI explicitly notes that research/API evaluations can differ from production behavior. The benchmark numbers therefore should not be treated as universal real-world performance guarantees.

2. GPT-6.1 Astra did not become the DevDay model

One notable part of the event was also what OpenAI did not launch.

GPT-6.1 Astra was not among the DevDay launches. TechCrunch reported that OpenAI had dropped the planned release after safety concerns emerged during internal testing, while AP also reported that OpenAI paused the model's release amid safety scrutiny.

Instead, OpenAI introduced GPT-6.1 Sol as the newer model in the GPT-6.1 family.

That matters because it shows the difference between the product roadmap and what OpenAI is actually willing to deploy.

OpenAI's published safety material for Sol classifies it as Critical in cybersecurity capability and High for biological and chemical capability, and says the model uses the same safeguards stack described for GPT-6 Astra.

3. Dots are OpenAI's new always-on AI agents

If GPT-6.1 Sol was the biggest model announcement, Dots was arguably the biggest product announcement.

OpenAI describes Dots as always-on agents designed to work continuously toward user-defined goals.

Each Dot gets its own cloud computer and browser. Dots can also connect to the user's approved applications through OpenAI's plugin ecosystem, which OpenAI says covers more than 4,000 apps.

The difference from normal ChatGPT is the persistence.

A normal chatbot waits for the next prompt.

A Dot can keep working after the user has moved on.

OpenAI's examples include:

• Watching customer feedback for recurring issues

• Preparing and testing software fixes

• Updating product-launch materials

• Re-running analysis when new research data arrives

• Keeping sales proposals aligned with changing requirements

• Preparing invoices and other recurring work

OpenAI says Dots can work through ChatGPT, Slack and Teams, with text messaging planned for a future update.

4. Dots have permissions, monitoring and approvals

OpenAI isn't presenting Dots as unrestricted autonomous workers.

Each Dot operates on its own cloud computer, while the user's own computer remains separate unless the user explicitly connects it.

Users choose which applications the Dot can access.

Dots also have built-in rules for when they can act independently and when they should ask for approval. Custom Rules can allow, require approval for, or block specific actions.

OpenAI also introduced auto-review for Dot actions that could affect accounts or share information.

The company says some sensitive actions, such as password changes, remain with the user.

That architecture is important because persistent agents create a much larger security surface than ordinary chat.

5. Specialist Dots target enterprise workflows

OpenAI is also testing Specialist Dots for organizations.

Unlike a personal Dot, a specialist Dot can have its own identity, credentials and access to specific systems of record. OpenAI describes possible responsibilities in areas such as procurement, invoice processing, email marketing, customer support and commercial contracting.

OpenAI is also working with Microsoft on integrating Specialist Dots with Microsoft Agent 365, allowing companies to manage those agents using Microsoft's governance and security controls.

This is a sign that OpenAI sees agents becoming part of enterprise IT infrastructure rather than remaining personal assistants.

6. Ultrafast brings a new speed tier for GPT-6 Astra

OpenAI also announced Ultrafast, a premium service tier designed for applications where latency is especially important.

OpenAI says Ultrafast can deliver up to 8× faster token generation in Codex and up to 6× faster generation in the API compared with standard processing.

GPT-6 Astra Ultrafast is available through the API and in ChatGPT Work and Codex for the new Pro 500 and eligible enterprise plans.

The API pricing page lists Ultrafast at six times the standard Astra price.

For short-context GPT-6 Astra, that means:

Input:         $60 / 1M tokens
Cached input:  $6 / 1M
Cache write:   $75 / 1M
Output:        $300 / 1M

The trade-off is straightforward: faster responses in exchange for substantially higher cost.

7. Agents API gets computer use

OpenAI's Agents API had already entered public beta earlier in September.

At DevDay, the company added computer use to it.

That lets developers build agents capable of interacting with software through its user interface rather than relying only on APIs.

The Agents API also incorporates capabilities associated with Codex, including:

• Tool calling

• Tool search

• Multi-agent orchestration

• Context compaction

• Hosted execution

• Long-running agent sessions

OpenAI says its infrastructure handles the underlying agent loop, while developers control what the agent can access and what environment it uses.

This is one of the most important pieces for developers because it takes functionality that OpenAI learned from Codex and makes it available as a general-purpose platform.

8. Decisions API turns Luna into a fast decision engine

Another new API is the Decisions API.

The concept is intentionally narrow.

Instead of asking a model to generate an unrestricted answer, developers define a finite set of allowed answers.

The model then selects one.

OpenAI says developers can provide context through text or images and use the result to:

• Classify content

• Route requests

• Select the next agent action

• Make other fast machine-readable decisions

The API is currently in limited preview, with broader release planned.

This could be useful for software systems where the application needs a predictable decision rather than a paragraph of generated text.

9. OpenAI and Amazon bring managed agents to AWS

OpenAI also announced Bedrock Managed Agents, powered by OpenAI, through its partnership with Amazon.

The service brings core Agents API capabilities into AWS so that developers can build OpenAI-powered agents that run within Amazon's cloud environment and interact natively with AWS resources.

For enterprises already committed to AWS, this reduces one of the practical barriers to adopting agentic applications.

The direction is important:

OpenAI models → OpenAI agent infrastructure → enterprise cloud environment

rather than forcing every enterprise to build and manage the entire agent execution layer itself.

10. Private Intelligence targets enterprise data protection

OpenAI introduced Private Intelligence for organizations that want stronger controls over sensitive data.

The company says Zero Data Retention with Private Safety Processing can perform automated safety reviews without giving OpenAI personnel access to the underlying customer content.

OpenAI also announced a preview of Private Inference, planned for later this fall.

The company describes it as a combination of confidential computing and strict, verifiable controls.

For companies working with financial, legal, healthcare or other sensitive information, this part of the event may be as important as the model releases themselves.

11. Codex is moving beyond the laptop

OpenAI dramatically expanded Codex's role at DevDay.

The company introduced Codex Cloud with reusable development environments.

Instead of treating every remote coding task as a separate disposable sandbox, developers can now prepare a reusable environment containing repositories, dependencies, tools and approved settings.

The result is a much more practical remote development workflow.

Developers can start a task on a computer, continue it from a phone and let it keep running in the cloud even while their laptop is closed.

OpenAI says Codex Cloud is available across Plus, Pro, Business, Healthcare, Education and Enterprise plans.

12. Codex CLI gets voice controls and an agents view

The refreshed Codex CLI adds two-way voice interaction.

OpenAI also introduced a new /agents view for managing and tracking multiple tasks.

Other improvements include:

• Better prompt editing

• Session resumption

• Worktree workflows

• A cleaner terminal interface

These changes are less flashy than Dots, but they could matter more to developers who spend hours inside the terminal.

13. Code Review becomes part of the Codex workflow

OpenAI introduced a new Code Review experience inside the ChatGPT desktop app.

Developers can inspect summaries and diffs and ask Codex questions about potential issues before sending feedback on GitHub pull requests or GitLab merge requests.

OpenAI also says automatic reviews can run in the cloud.

That moves Codex closer to a full software-development workflow:

Idea
↓
Implementation
↓
Tests
↓
Code Review
↓
Security Scan
↓
Human approval

14. Codex Security Cloud turns Codex into a security tool

One of the most practical DevDay launches for software teams is Codex Security Cloud.

It can scan entire GitHub repositories on demand or on a schedule, monitor new commits, investigate findings, remove duplicate findings and prepare fixes in the cloud.

OpenAI also tied the product to its Daybreak Blue security access program.

This means Codex isn't being positioned only as a coding assistant anymore.

It is becoming part of the broader software-security lifecycle.

15. OpenAI is turning ChatGPT into a developer platform

Another major theme of DevDay was the evolution of plugins.

OpenAI introduced Plugin Extensions, giving developers a way to create more native experiences inside ChatGPT.

Extensions can:

• Add an entry to the ChatGPT sidebar

• Create interactive panels

• Provide file viewers

• Bring app functionality into conversations

OpenAI specifically mentioned Figma and Adobe integrations during the event.

This is a significant shift from a simple “connect my app to ChatGPT” model.

The application can increasingly become part of the ChatGPT interface itself.

16. Plugin creation and discovery are getting an overhaul

OpenAI also updated the developer experience around plugins.

The company says Plugin Creator is being improved, the submission process now provides clearer feedback and review tracking, and discovery is being improved through better ranking and recommendations.

The practical objective is to make the plugin ecosystem easier to build for and easier for users to navigate.

That becomes especially important if ChatGPT is going to operate as a software platform rather than simply an AI assistant.

17. Sites can now host plugins

OpenAI announced that supported plugins can now be added to Sites built within ChatGPT.

Users in a workspace can use the same app while maintaining their own connected data and permissions.

This creates a new development pattern:

Website/Site + ChatGPT + plugins + personalized user data

Instead of building every integration separately into a traditional web app, some functionality can be delegated to ChatGPT's connected-app system.

18. MCP Events can trigger automations

OpenAI also added support for the proposed MCP Events specification.

This allows plugins to react to events from connected apps.

For example, if a new task arrives in a project-management tool, ChatGPT can automatically read the relevant documents and prepare a plan.

This is especially important for automation.

The old model is:

User asks → AI responds

The new model is:

Event happens → AI notices → AI takes an authorized action

That is a fundamental change in how AI applications can operate.

19. ChatGPT Space creates a shared home for people and agents

OpenAI also introduced ChatGPT Space, a collaborative environment for teams.

Space gives teammates, ChatGPT and their Dots a shared place for files, project context and ongoing work.

Instead of keeping work spread across separate chats and folders, a Space can act as a common project environment.

OpenAI says Space is available on desktop and web for Pro, Business and Enterprise users, with mobile support for finding, reading and sharing pages.

20. Pages are editable documents for humans and agents

Inside those spaces, OpenAI introduced Pages.

Pages are designed as editable documents where people and AI can work together.

OpenAI says users can write, research, generate charts, create images and visualize information directly within a Page, while teammates can contribute ideas and feedback.

This turns ChatGPT into something closer to a collaborative document environment.

It also creates a new interaction model:

Prompt → document → AI edits → human feedback → AI revision

rather than simply:

Prompt → answer

21. Collaborative slides are coming next

OpenAI also showed collaborative slides.

The feature will allow multiple teammates and agents to work on interactive presentations from conversations or templates.

OpenAI says users will be able to present inside ChatGPT or export to PowerPoint and Google Slides while preserving formatting.

The feature is planned for the coming weeks rather than being fully available on DevDay day one.

22. Teams and Team Tasks bring recurring work into ChatGPT

OpenAI also introduced team collaboration features that allow colleagues to share pages, slides, plugins and spreadsheets.

More importantly, Team Tasks lets teams delegate recurring work.

OpenAI says ChatGPT can use connected tools to gather information and act on a schedule or in response to events such as a new email or Slack message.

This is very close to the kind of workflow automation businesses have historically built using specialized automation platforms.

23. ChatGPT is moving directly into Slack and Microsoft Teams

OpenAI announced deeper ChatGPT integration with Slack and Microsoft Teams.

Team members can mention @ChatGPT in channels, threads or direct messages and get assistance without leaving the conversation.

OpenAI says ChatGPT can use tools connected by the workspace administrator or tools connected by the individual user, subject to permissions.

That matters because enterprise AI adoption is often determined by whether employees have to switch applications.

Here, the AI comes into the application where the conversation is already happening.

24. Meetings plugin turns conversations into action items

OpenAI also announced a Meetings plugin.

It can create notes and action items and save personalized summaries into ChatGPT Space.

OpenAI says meeting audio is deleted after the notes are created and cannot later be replayed through the plugin.

The initial beta is available on macOS for Pro and Business users, with Enterprise support planned.

25. Sign in with ChatGPT could change how AI apps are paid for

One of the most commercially interesting announcements was Sign in with ChatGPT.

OpenAI is allowing users to sign into participating third-party apps using their ChatGPT account.

More unusually, eligible Plus and Pro users can also use their existing OpenAI subscription allowance across participating tools. OpenAI says there are 16 launch partners.

The initial ecosystem includes tools such as:

• Devin

• Notion

• Vercel

• OpenClaw

• Dactyl

• Warp

• Kilo Code

• Conductor

This creates a different business model for AI applications.

Instead of every application buying tokens from OpenAI and charging customers separately, some users can use part of their existing ChatGPT allowance.

That could make building AI products easier for developers and cheaper for users, although usage limits and eligibility vary by product and plan.

26. OpenAI Marketplace targets enterprise software

OpenAI also announced the OpenAI Marketplace in beta for eligible enterprise customers.

The idea is that companies can apply part of an existing OpenAI commitment toward approved partner software.

OpenAI says its first 32 partners include companies such as:

This takes OpenAI further into enterprise software distribution.

27. Pro 500 is the new high-usage ChatGPT tier

OpenAI also launched Pro 500, a new $500-per-month tier.

The plan is designed for users who want the highest usage allowance and access to Ultrafast.

OpenAI says Pro 500 provides 25× the Plus usage allowance.

The pricing is high, but the target audience is equally specific: developers, researchers and professionals who regularly run long, resource-intensive AI workflows.

OpenAI also reopened Pro 200 subscriptions, with allowances that differ for new users versus some existing subscribers.

28. Shareable profiles make ChatGPT creations easier to distribute

OpenAI also introduced Shareable Profiles.

Users can collect their Sites and plugins into one page that other people can discover and reuse.

OpenAI says workspace members can also discover shared skills, while skills sharing is restricted to the workspace.

This is another small but important step toward making ChatGPT a platform with creators, developers and reusable AI experiences.

What OpenAI was really saying at DevDay

When all of the announcements are put together, a much clearer strategy appears.

The core idea can be simplified into three layers:

That strategy turns ChatGPT into something much larger than a chatbot.

It becomes a platform connecting:

Users → AI → agents → applications → tools → enterprise systems

TechCrunch similarly interpreted the announcements as a move toward making ChatGPT a place where software can be discovered, launched and used directly.

Why Codex may be the most important developer story

Dots grabbed most of the attention, but Codex may have the more immediate impact on professional developers.

OpenAI is turning Codex into a broader software-delivery environment:

Write code
   ↓
Run remotely
   ↓
Continue in the cloud
   ↓
Review changes
   ↓
Scan security
   ↓
Prepare fixes
   ↓
Human approves

That workflow spans Codex Cloud, Code Review, Codex Security Cloud and the Agents API.

This is a much bigger ambition than code autocomplete.

The system is becoming an AI-managed development environment.

Why the Agents API matters even more

For developers building their own applications, the Agents API is perhaps the most foundational release.

OpenAI is effectively exposing the machinery it built for Codex:

• Long-running sessions

• Tool orchestration

• Context management

• Subagents

• Computer interaction

• Sandboxed execution

That means companies don't necessarily have to build the full infrastructure required for reliable agent execution from scratch.

The real competition may therefore shift from:

Who has the best chatbot?

to:

Who can build the most useful agent around a strong model?

OpenAI is also making ChatGPT more like an operating layer

Several DevDay announcements point toward the same endpoint.

Plugin extensions allow software to live inside ChatGPT.

MCP Events let external events trigger AI workflows.

Sign in with ChatGPT connects identity and subscription usage to external apps.

Marketplace connects enterprise spending to partner software.

Space and Pages make ChatGPT a shared workspace.

Dots allow persistent agents to operate inside that workspace.

Put together, these components start to resemble an AI operating layer.

Not an operating system in the traditional Windows or Linux sense.

Rather, a layer sitting between users and software.

The safety problem gets bigger as autonomy gets bigger

There is a clear trade-off running through the entire event.

More powerful agents can accomplish more.

But agents with access to browsers, files, code, accounts and external applications can also make more consequential mistakes.

OpenAI's recent safety work reflects that.

The GPT-6.1 Sol safety documentation describes stronger safeguards, while Dots include permissions, monitoring and action review. Codex Security Cloud brings security scanning directly into the agent workflow.

That suggests OpenAI increasingly sees security and agent infrastructure as the same problem.

What developers should take away from DevDay 2026

For developers, the event changes the toolkit considerably.

A modern OpenAI-based application can now include:

• GPT-6.1 Sol for lower-cost complex work

• GPT-6 Astra when maximum capability matters

• Agents API for autonomous workflows

• Computer use for software interaction

• Hosted sandboxes for execution

• Decisions API for constrained decisions

• Codex Cloud for persistent development environments

• Code Review for automated review

• Codex Security Cloud for security workflows

• Plugins and extensions for ChatGPT-native experiences

• MCP events for event-driven automation

That is a substantially larger developer platform than OpenAI had only a year earlier.

What businesses should take away

For businesses, the event was less about individual model benchmarks and more about infrastructure.

OpenAI is now offering mechanisms for:

Employee
   ↓
ChatGPT
   ↓
Agent
   ↓
Company tools
   ↓
Data
   ↓
Action
   ↓
Audit / approval

That workflow is increasingly becoming the core unit of enterprise AI.

The difficult part isn't simply model intelligence.

It is permissions, governance, identity, security and accountability.

OpenAI's announcements around Private Intelligence, Specialist Dots, Microsoft Agent 365 and Marketplace show that enterprise deployment is becoming a central part of the company's strategy.

The biggest unanswered question

OpenAI has now assembled most of the pieces required for persistent AI agents.

The question is no longer whether an AI can generate a useful answer.

The question is whether people will trust it enough to keep working without supervision.

Dots are designed around exactly that proposition.

They can work in the background.

They can access connected tools.

They can maintain context.

They can take action.

They can ask for approval when needed.

That is a fundamentally different relationship between humans and software.

The bottom line

OpenAI DevDay 2026 was not a normal developer conference built around one flagship model.

It was a platform expansion.

OpenAI introduced GPT-6.1 Sol, Dots, Ultrafast, Agents API computer use, Decisions API, Codex Cloud, a refreshed Codex CLI, Code Review, Codex Security Cloud, plugin extensions, MCP Events, Space, Pages, collaborative slides, Team Tasks, Slack and Teams integration, Meetings, Sign in with ChatGPT, Pro 500 and the OpenAI Marketplace, alongside enterprise privacy and governance updates.

The common thread is simple:

AI is moving from answering prompts to participating in workflows.

GPT-6.1 Sol makes capable agentic intelligence cheaper.

Agents API gives developers the infrastructure to build agents.

Codex gives those agents a software-development environment.

Dots make agents persistent.

Plugins and MCP let them connect to applications.

Space and Pages give humans and agents a shared workspace.

Marketplace and Sign in with ChatGPT extend the ecosystem beyond OpenAI's own interface.

The result is a much bigger ambition than “build a better chatbot.”

OpenAI is building toward a world where the AI becomes the layer through which people interact with software, while agents continue working after the conversation ends.