OpenAI, the company behind ChatGPT, held DevDay, its developer conference, on September 29, 2026. The announcements covered AI agents that keep working between conversations, tools for building software, and shared spaces where people work with AI.

OpenAI's official recap connects the launches to agents taking on ongoing responsibilities and people collaborating with AI. My practical reading: OpenAI wants to give AI a continuing role in getting work done. That matters whether you are choosing software for a team or building a product of your own.

This recap explains the launches, who can use them, and which features remain in preview or are coming later. Availability was checked on October 5. Related October updates appear separately below.

Dots is OpenAI's new persistent-agent offering. A dot has its own cloud computer, connects to your tools, and can continue working between conversations. That makes the product relevant to recurring research, preparation, and coordination rather than only questions you ask in an open chat.

Access is rolling out to Pro and Business Premium users in eligible markets. Enterprise, Edu, and Healthcare access is an admin-enabled beta. The specialist enterprise dots shown alongside the launch are pilots. Those are different stages of availability, and a demonstration does not tell you which version your organization can use.

My first question would be what work deserves that persistence. A changing project brief is a more useful starting point than an instruction to watch everything. That question puts Dots in context within a much wider announcement list.

Read the Dots announcement →

OpenAI introduced an updated Sol for coding, computer interaction, and professional work. The GPT-6.1 Sol announcement lists standard input and output token prices at one-fifth of Astra's. It is available through the API and in Work and Codex for eligible plans, but not yet in ordinary Chat. OpenAI's capability comparisons are vendor evaluations; the pricing ratio does not establish a matching reduction in the cost of a completed task.

Ultrafast addresses speed separately. Astra is available through this premium API tier, subject to rate limits. Work and Codex access is available on Pro 500 and eligible Enterprise and Edu plans, with workspace restrictions. Sol Ultrafast is still coming soon. Faster token generation can help an interactive workflow, but a tool call or a slow external application may still dominate the wait.

These announcements give buyers two questions to ask: which model produces acceptable work, and when is faster generation worth paying for? One answer need not determine the other.

Cloud execution. Codex Cloud runs tasks in isolated workspaces based on reusable development environments. A task can continue after the laptop closes. A prepared environment is useful for work that repeatedly needs the same repository, dependencies, and checks. It still needs to resemble the environment in which the proposed change must work.

The terminal. The refreshed CLI adds voice steering, an agents view, and improved session and worktree handling. These are changes to how developers direct and follow work.

Code Review. The desktop review experience puts changes, comments, and checks together. The current documentation distinguishes generally available GitHub support from GitLab merge-request support in preview. A clearer review surface is useful; a team still decides which checks and approvals a change needs.

Security Cloud. Codex Security Cloud scans connected GitHub repositories and follows new commits for security findings. It adds a recurring security job to the coding workflow. Repository access and workspace permissions remain part of setup. Findings and proposed fixes deserve investigation before acceptance.

Taken together, these changes cover preparation, execution, review, and security maintenance. A development team can evaluate each against a bottleneck it already has.

The Agents API computer-use update lets an agent operate an OpenAI-hosted browser. Developers receive session events and handle access requests and sign-in. This can reduce the browser infrastructure a product team must operate. The application still needs a useful response when access is denied or a task cannot finish.

The Decisions API uses Luna to answer predefined questions with finite answer choices, supporting classification and routing. It was announced in limited preview. A planned broader release is not confirmation of account access.

AWS Bedrock Managed Agents, powered by OpenAI, provides another deployment route, combining OpenAI capabilities with AWS services. AWS labels it preview. For organizations already operating in AWS, that is a concrete option to investigate, subject to its preview limits.

These are different building blocks. A hosted browser is relevant when software interaction is the obstacle. A bounded decision interface is relevant when the next step depends on a small set of answers. The cloud route matters when deployment requirements determine what a team can adopt.

Plugin Extensions can give a product a sidebar presence, interactive panels, or file viewers inside ChatGPT. The detailed documentation is narrower than the launch shorthand: web extensions for Free and Go are coming soon, and composer mentions are desktop-only. Check the actual surface your intended users have.

OpenAI also announced Plugin Creator, an improved submission flow, and changes to plugin discovery, as described in the DevDay recap. That is relevant to the work of getting an integration built and found.

ChatGPT Sites can now use supported plugins in workspace applications. Each visitor uses their own connections and permissions. A shared interface therefore does not automatically share its creator's access to outside systems. This matters for a team dashboard whose users have different access rights.

MCP Events lets connected-app events trigger a defined workflow. In plain English, work can start when something changes in another system. Developers and workspace administrators must support the event connection and its controls. A notification becoming a work trigger is a meaningful change, but the trigger still needs a specific assignment.

For a product team, the useful question is where customers should encounter its software: a standalone site, a shared internal application, or a panel beside their AI conversation.

Space becomes the shared home for files and work with teammates and AI. Pages are editable documents that bring text, images, charts, and interactive material together. Pro, Business, and Enterprise users can create and edit on web and desktop. Mobile supports finding, reading, and sharing; creation and editing are coming later. Collaborative slides are also coming soon. A reader should not mistake the product demonstration for availability of every format.

Teams and Team Tasks organize shared work and recurring assignments. Administration, service accounts, and connections determine how the work reaches external systems. The practical benefit to examine is whether a recurring report or update can have a shared purpose and a visible result.

ChatGPT in Slack and Microsoft Teams brings assistance into channel discussions. The current product page lists Business, Enterprise, and Edu with admin setup. For teams, the point is getting help where a discussion is already happening.

The Meetings plugin produces personalized notes and action items. It is a macOS beta for Pro and Business; Enterprise is in limited alpha. Other operating systems are coming later. Meeting capture and the resulting notes need to fit the team's actual meeting practices.

Shareable profiles give people a way to present themselves and selected creations. Current Help documentation says Enterprise support is coming soon and the Sites showcase is unavailable on mobile. Again, the detailed rollout matters more than a blanket claim that every plan has everything.

The attraction here is continuity: a discussion, its supporting material, and the resulting document can become easier to find together. Whether that reduces handoff work is something each team can test on a real project.

Sign in with ChatGPT has two separate functions. People can use their identity in participating applications. Eligible Plus and Pro users can also choose to use their plan allowance for eligible AI requests in those tools. App-level limits do not create extra allowance, and third-party subscriptions can still cost separately.

The new Pro 500 tier adds a higher allowance and Astra Ultrafast access. It is another plan choice, not evidence that every user needs a larger subscription. Match the tier to work that actually runs into a useful limit.

OpenAI Marketplace lets eligible enterprise customers request to use part of an existing OpenAI commitment toward approved partner products. The entry point is a purchase request or expression of interest. This is an enterprise purchasing arrangement, not a general perk attached to every personal subscription.

Under Private Intelligence, Private Safety Processing supports automated safety review for organizations approved for Zero Data Retention. It has configuration requirements. Separately, the Private Inference preview is planned for the fall. One should not be treated as proof that the other is available today.

These details belong with a feature evaluation. The right product can still be the wrong fit for a particular account, data requirement, or procurement arrangement.

On October 5, OpenAI announced a visual advertising format, expanded measurement, and brand-suitability work. The visual format is scheduled to begin testing later this month in the US with selected advertisers, initially during image generation. OpenAI says ads will be labeled and separate from the generated image, and that advertising does not influence ChatGPT's answers.

This October 5 announcement followed the September 29 DevDay launches. For marketers, it is a channel development to follow and evaluate as access opens. It does not yet provide grounds to assume that a campaign will beat an existing channel.

A model guide for the GPT-6 family

OpenAI's October 2 guide follows the launches with advice on model selection, instructions, long-running work, and evaluation. Read it for the practical discipline of measuring cost per successful task, rather than treating a token price as the whole budget.

Read the guide →

My takeaway from the full list is to keep three columns in your notes: what was announced, what your account can use, and what would improve work you already do. The announcements are broad enough that nobody needs to test everything this week. A complete map should help you choose where to begin.