Everything OpenAI announced at the DevDay 2026 keynote

OpenAI announced more than 20 major updates at DevDay in San Francisco on September 29th. The keynote covered always-on agents, model pricing, cloud coding and distribution. Some products are available today; others are entering previews or rolling out over the coming weeks.

Here is the rundown, with availability checked against OpenAI's launch materials and product documentation.

Dots: always-on agents with their own computers

Dots are persistent agents powered by GPT-6 Astra. Each gets a cloud computer and can work across connected apps, carry context between conversations and handle several ongoing projects.

Users can reach their dot through ChatGPT, Slack and Microsoft Teams, including voice calls. Texting is coming later. With permission, a dot can also connect to a user's laptop.

OpenAI demonstrated delegating Slack requests, investigating bugs and preparing pull requests. Its background “proactive research” uses read-only connected tools; actions follow permissions and approval rules.

OpenAI's launch materials say dots are rolling out to Pro and Business Premium users in eligible markets. Enterprise, Edu and Healthcare workspaces can enable the beta through an administrator. A first dot is included without an additional subscription charge, although Work and Codex tasks still consume the usual plan allowance.

Specialist dots and Microsoft Agent 365

OpenAI also previewed specialist dots for organizational responsibilities, with their own identities, credentials and system access. Initial enterprise pilots cover defined workflows.

The company is working with Microsoft to integrate them with Agent 365's governance and security controls.

GPT-6.1 Sol: lower-cost intelligence for coding and work

GPT-6.1 Sol upgrades GPT-6 Sol. OpenAI says it approaches Astra's performance on coding, computer use and professional tasks at one-fifth of Astra's standard input and output token prices.

OpenAI's published API pricing is:

The 95% discount discussed onstage applies to cached input relative to standard input pricing. Reusing context therefore costs substantially less than sending fresh input.

The model is available through the API as gpt-6.1-sol and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. OpenAI says it is not yet available in ordinary Chat.

Astra Ultrafast: a premium speed tier

OpenAI introduced Ultrafast, a faster inference tier for GPT-6 Astra, following the eight-times speed announcement RuntimeWire covered during the keynote. The keynote described speeds of 300 tokens per second and pricing at six times the standard rate.

OpenAI's launch recap specifies up to eight times faster token generation in Codex and up to six times faster in the API. Those figures describe generation speed; a task's overall completion time also depends on tools and other work.

Astra Ultrafast launches today in the API and in ChatGPT Work and Codex on Pro 500 and Enterprise plans. GPT-6.1 Sol Ultrafast is coming soon.

RuntimeWire previously covered the August GPT-5.6 Sol Ultrafast preview on Cerebras hardware, which carried a different, up-to-14-times speed claim. That earlier preview concerned a different model and rollout.

Pro 500 launches, and Pro 200 reopens

The new Pro 500 subscription costs $500 a month, includes Astra Ultrafast and offers an allowance OpenAI describes as 25 times Plus usage. Eligible usage can also be spent in participating partner apps.

OpenAI reopened the $200-a-month Pro tier to new subscribers. The updated terms matter: new subscriptions without grandfathering receive a lower allowance than the previous Pro 200 offering. Eligible existing subscribers retain their earlier allowance through October 29th, 2026, before moving to the reduced allowance at the same monthly price.

Decisions API: fast answers from a fixed set of choices

The Decisions API focuses Luna on questions with predefined possible answers. Developers supply text or images and receive decisions they can use for classification, request routing or an agent's next action.

Onstage, OpenAI described responses arriving in a fraction of a second and discussed using vision inputs to help hardware react quickly.

It enters limited preview today, with broader availability planned in the coming days.

Agents API: OpenAI's managed Codex harness

The new Agents API gives developers access to the Codex harness through a managed service. OpenAI handles sessions, orchestration, context compaction and recovery; developers supply tools and choose execution environments.

Agents can execute code, edit files, connect to MCP servers and delegate work to other agents. OpenAI presented the API as a public beta.

Computer use in the Agents API

Computer use lets agents navigate websites and operate applications through an OpenAI-hosted browser. Developers can follow the session's events while the agent observes the interface and decides what to do next.

OpenAI also listed computer use in ChatGPT Work and Codex for Pro 500 and Enterprise.

Bedrock Managed Agents: OpenAI agents running inside AWS

OpenAI and Amazon introduced Bedrock Managed Agents, combining OpenAI's harness with AWS infrastructure and integrations.

Amazon says agents run in the customer's environment, inference runs on Bedrock and data stays within AWS. The service handles inference, memory and skills and works with Bedrock AgentCore.

Codex in the cloud gets reusable environments

Codex Cloud lets developers run tasks remotely and continue from desktop, web or mobile.

Reusable development environments hold a project's repositories, dependencies, tools and access settings. Each task gets its own workspace, allowing teams to reuse an approved setup while work continues with a laptop asleep.

RuntimeWire identified the new Codex Cloud path in desktop build 9922 on September 20th, nine days before the keynote. The investigation traced agent-led environment setup alongside Tailscale, proxy-delivered secrets and cloud-identity controls. Those findings described gated client code; the keynote announced reusable cloud environments.

The Codex CLI gets voice and a new agents view

The refreshed Codex CLI lets users start and steer tasks by voice.

A new /agents view helps track delegated work and multiple tasks. OpenAI also announced improvements to prompt editing, session resumption, Git worktrees and the terminal interface.

Code review moves into the desktop app

A new Code Review experience in the ChatGPT desktop app brings pull request summaries, changed files, comments and checks into one interface.

Users can inspect diffs and ask Codex to investigate issues before posting feedback. Automatic cloud reviews can run while a developer is away. GitHub support is generally available; GitLab merge request support is in preview.

Codex Security Cloud scans repositories and prepares fixes

Codex Security Cloud scans connected GitHub repositories on demand or on a schedule and monitors new commits.

OpenAI says it investigates findings, removes duplicates and prepares fixes in the cloud. It also includes access to models offered through Daybreak Blue without a separate Daybreak application.

Private Intelligence adds protected safety processing

OpenAI announced Private Intelligence for businesses handling sensitive data.

Zero Data Retention with Private Safety Processing supports automated safety reviews without human access to protected content. Encrypted safety records are kept in customer-controlled storage and reviewed in a hardware-attested runtime.

OpenAI separately previewed Private Inference, coming this fall, which combines confidential computing with verifiable controls.

Plugin extensions let developers build fuller apps inside ChatGPT

Plugin extensions let developers add sidebar destinations, interactive panels and file viewers to ChatGPT.

The keynote showed Figma designs and comments alongside a conversation, Photoshop functionality inside ChatGPT, and OpenAI's own Meetings app.

Developers can build editors, dashboards and workspaces using the same platform OpenAI uses for its features. Documentation notes that web extensions for Free and Go users are coming soon; composer mentions currently require the desktop app.

Plugin creation, submission and discovery are changing

Plugin Creator helps users describe a workflow, add instructions and reference files, then build and refine a reusable plugin through conversation.

OpenAI also announced a simpler submission process: developers can track reviews, see what needs fixing, request human review and update tools without restarting the entire submission.

Plugins can be recommended within conversations when ChatGPT identifies a relevant capability, alongside discovery through the directory.

ChatGPT Sites can host connected plugins

Sites can now use supported ChatGPT plugins and connected data. Team members can use the same app with their own accounts, data access and permissions.

OpenAI is also improving how developers add and manage automations so Sites can keep shared information current.

MCP events can trigger automations

Support for the proposed MCP Events specification lets connected apps notify ChatGPT when something changes.

A new task, message or content update can trigger an automation while the user is away. Users choose what ChatGPT should monitor and what it should do when an event arrives.

The integration uses webhooks and requires developer implementation; connecting an app alone does not establish an automation.

ChatGPT Space creates a shared home for team work

ChatGPT Space brings teammates, ChatGPT and dots together around shared knowledge and documents.

The keynote demonstrated tagging colleagues and agents, building interactive visualizations and using connected Slack context to update work. Space is available on Pro, Business and Enterprise, with desktop and web support.

Pages

Pages are a new document format for collaboration between people and agents. They can contain writing, research, charts, images and interactive work, with comments and real-time editing.

Connected tools can help keep pages updated.

RuntimeWire's August 28th investigation found a gated document-review system that turned selected Excel cells and PowerPoint elements into assignments for ChatGPT. That earlier reporting documented threaded agent replies and edits to the underlying Office files.

Collaborative slides

OpenAI previewed slides that teammates and agents can edit together, including comments and support for starting from a conversation or an existing template.

Users will be able to present in ChatGPT or export to PowerPoint and Google Slides. Availability on Pro, Business and Enterprise is planned for the coming weeks. The Space product page also lists collaborative spreadsheets as coming soon.

The presentation tools extend OpenAI's push into document creation. RuntimeWire covered ChatGPT's September 17th arrival in Microsoft Word, which completed its add-in lineup across Word, Excel and PowerPoint.

Teams and shared tasks support recurring work

Teams and Team Tasks let coworkers organize shared work and jointly manage recurring instructions.

Tasks can run on a schedule or respond to events such as an email or Slack message. Company-managed connections let teams use shared external accounts.

ChatGPT can join Slack and Microsoft Teams conversations

Users can mention @ChatGPT in Slack or Microsoft Teams to request work in a channel, thread or direct message.

ChatGPT can use authorized connected tools, and teammates can contribute context and refine the result within the conversation. OpenAI says colleagues can participate without each needing an individual ChatGPT license.

Meetings turns conversations into notes and follow-up work

The Meetings plugin captures online or in-person meetings and saves personalized summaries and action items in Space.

Users can connect a calendar, share notes and ask ChatGPT to perform follow-up work. Audio is deleted once notes are ready.

The beta launches in the macOS desktop app for Pro and Business. Enterprise is in limited testing, with wider support coming later; Windows, iOS and Android support is also planned.

Shareable profiles showcase what users build

Shareable profiles let users display their Sites and activity through a profile that can be kept private or shared.

The help center says profiles are rolling out gradually and Enterprise support is coming soon.

Sign in with ChatGPT lets subscriptions travel to other apps

Sign in with ChatGPT lets Plus and Pro customers use their existing Work and Codex allowance in participating products.

OpenAI announced 16 participating tools, including Devin, Notion, Vercel, T3, OpenClaw and Dactyl. Users can set per-app limits, and eligible requests count against their overall plan allowance.

The partner's own subscription or infrastructure charges may still apply. Identity sign-in and access to ChatGPT plan usage are separate capabilities: some products support the login without token sharing.

OpenAI Marketplace lets enterprises buy partner software

OpenAI Marketplace lets eligible enterprise customers apply part of their existing OpenAI spending commitment toward approved partner products.

OpenAI announced 32 initial partners, including Adobe, Figma, Sierra, Decagon, Salesforce, ServiceNow, Harvey, Legora, Palo Alto Networks, CrowdStrike and Baseten.

For participating vendors, that creates a sales channel tied to budgets customers have already committed to OpenAI. Enterprise customers can express interest, and vendors can join a partner waitlist.

An AI research intern and faster computer use

OpenAI repeated its claim that it had reached its September goal of an automated research intern: a system that can take on a clearly defined task requiring substantial effort from a skilled researcher. As RuntimeWire's research-intern coverage explains, the milestone had already appeared in OpenAI's September 6th research report; the keynote restated it.

The company said its models can handle more than a third of day-long research tasks. Researchers also described using models to optimize the computer-use harness and improve safety monitoring. Those were OpenAI's keynote claims, rather than independently verified results.

Reporting draws on RuntimeWire's live keynote coverage, OpenAI's DevDay recap and the product documentation linked above. Availability reflects September 29th release information and varies by plan, market and rollout stage.