OpenAI DevDay 2026: what each announcement means for enterprises

A dot is an always-on agent powered by GPT-6 Astra, with its own cloud computer and browser. It learns from feedback, works towards the goals you set around the clock, connects to more than 4,000 apps through plugins and can run several projects at once. You can open its computer to see what it is doing, and with your permission it can use your laptop.

You reach the same dot in ChatGPT on desktop, web and mobile, in Slack and in Microsoft Teams, with texting to follow, and OpenAI says it carries its context across all of them. The first dot is included in Pro and Business Premium at no extra cost, with an allowance for deeper work and extended limits in the first month. Talking to it does not count against ChatGPT usage limits; tasks it starts in Codex or ChatGPT Work do.

Rollout began on 29 September 2026 for Pro and Business Premium in eligible markets. On Enterprise, Edu and Healthcare it is a beta, off until a workspace admin enables it. Specialist dots, with their own identity, credentials and IT-provisioned hardware, are a preview run as focused enterprise pilots with OpenAI's engineers, and integrate with Microsoft Agent 365 for governance and security.

Dots change how an agent sits inside a company. An assistant waits to be asked; a dot researches in the background with read-only tools, works to goals and turns up where work already happens. That makes it a new kind of account in your estate, with its own computer, saved passwords and plugins into your systems, acting for a person when nobody is watching.

OpenAI has shipped controls for that: Custom Rules to allow an action, require approval for it or block it, an Activity View, automatic review of actions, monitoring that can pause or stop a dot, and saved passwords the model never sees. Sensitive steps such as changing a password stay with the person. OpenAI does not train on Business, Enterprise or Edu content by default, and says it does not train directly on a dot's research or notes.

Cross-channel context is not unique to OpenAI: Meta says its consumer agent Muse, launched on 8 September, keeps context across the apps a user authorises. The difference is that dots land inside a business workspace with admin switches, which is the part an IT team can govern.

ChatGPT Space is a home for a team, where teammates, ChatGPT and your dot build on shared knowledge that ChatGPT keeps organised to your instructions. It is on Pro, Business and Enterprise on desktop and web. On mobile you can read and share now, with creation to follow.

Around it came Pages, documents built for people and agents to work on together; collaborative slides in the coming weeks; teams with shared tasks that run on a schedule or an event, on Business and Enterprise; a Meetings plugin in beta on the macOS app, which deletes the audio once the notes are ready; and shareable profiles.

Shared team knowledge usually lives in a separate wiki, which the assistant reaches through a connector if at all. Space puts it where the model and the agent already are, which removes one copy of the truth and one connector to maintain.

The fair comparison is Notion, whose Custom Agents already run on schedules or triggers while you are offline, free to try and then $10 per 1,000 monthly Notion credits. Notion is also one of the 16 partner tools in OpenAI's Sign in with ChatGPT list, so this is not a simple replacement story. If you are still choosing between vendors, our comparison of ChatGPT Work against Claude Cowork sets out where each fits.

Scheduled team tasks act without anyone pressing a button, so each needs an owner, a scope and a log, like any other automation.

Codex now runs on your computer, remotely from your phone, or in the cloud from any device. Teams get reusable development environments that carry shared, approved settings and permissions. It is on Plus, Pro, Business, Healthcare, Education and Enterprise.

The Codex CLI was refreshed on every plan: voice to start and steer tasks, an /agents view for running several tasks at once, better prompt editing, session resume and worktrees.

Running in the cloud changes where the work happens, and the security questions move with it. Code, secrets and network access now sit in an environment OpenAI runs, so the environment definition becomes the control: which repositories it can clone, what it can reach on the network and which credentials it holds.

Reusable environments with approved settings are the useful part for a platform team. Instead of every engineer configuring an agent on their own laptop, you publish one environment that carries the rules, and everyone starts from it. That is the operating layer we describe in our guide to rolling out coding agents across an engineering team.

Codex Security Cloud scans whole GitHub repositories on demand or on a schedule, keeps checking new commits, and investigates, deduplicates and prepares fixes in the cloud with your laptop closed. It includes access to the Daybreak Blue model and is on Pro, Business, Enterprise and Edu.

Code Review arrives in the ChatGPT desktop app on every plan: summaries and diffs, a way to ask Codex about an issue before you comment on a GitHub pull request or a GitLab merge request, and automatic reviews that run in the cloud while you are away.

Scheduled scanning with prepared fixes saves time where security work often stalls, which is triage. The risk is volume: a scanner that proposes fixes on every commit can bury a team in pull requests nobody has time to read. The deduplication OpenAI describes is aimed at that problem, and it is the first thing to test.

Automatic reviews raise a process question more than a technical one. Decide whether an AI review is advice or a gate, who may dismiss its findings and how a dismissal is recorded, before it runs on every merge request.

The Agents API lets developers build agents that operate software through computer use. It brings over Codex's multi-agent capabilities, tool search, tool calling and context compaction, and OpenAI runs the infrastructure. It is in the API, and in Codex and ChatGPT Work on Pro 500 and Enterprise.

Bedrock Managed Agents, powered by OpenAI, was built with Amazon: OpenAI agents that run entirely in AWS. A Decisions API also launched in limited preview, in which the Luna model answers questions that have a fixed set of answers, for jobs such as classifying a request, routing it or choosing an agent's next step.

Computer use matters for the systems that never got an API, such as an old finance tool or a supplier portal that only works in a browser. It is also the least predictable way to automate, because screens change and there is no contract to test against.

Bedrock Managed Agents suits organisations whose data, identity and spend already sit in AWS. Before choosing either, settle whether an agent is the right tool at all, or whether a plain integration would do the job.

GPT-6.1 Sol upgrades GPT-6 Sol, and OpenAI points to agentic coding, computer use and professional work. OpenAI positions it close to GPT-6 Astra in capability at a fifth of Astra's standard input and output token prices. It is in the API and on Plus, Pro, Business, Enterprise and Edu.

Ultrafast is a premium speed tier: up to eight times faster generation in Codex, where OpenAI quotes 300 tokens a second, and up to six times faster in the API. GPT-6 Astra Ultrafast is in the API, and in ChatGPT Work and Codex on Pro 500 and Enterprise. GPT-6.1 Sol Ultrafast is coming soon.

A model at a fifth of the flagship's token price is a cost decision before it is a capability one. Agent workloads are often long runs of routine steps, and those are where the cheaper model should be tried first, with the flagship kept for the steps that fail without it. Test that routing in your own evaluation, not a vendor's chart.

Speed is worth paying for where a person is waiting, such as a developer watching Codex or an agent in a live conversation. It is rarely worth it for work that runs overnight. Either way, set budgets and limits on AI spend per team before a faster tier makes spend grow faster too.

OpenAI grouped its privacy work under the name Private Intelligence. Zero Data Retention now comes with Private Safety Processing, which runs safety reviews automatically so that OpenAI staff do not access the content. Private Inference, which combines confidential computing with controls you can verify, is due in preview this autumn.

For regulated buyers, the open question with Zero Data Retention has been who can see content that a safety system flags. Private Safety Processing is OpenAI's answer: the review is automated and staff do not access the content.

Private Inference goes further: controls a security team can verify would let it check the claim rather than trust it. It is a preview with no date beyond this autumn, so it should not carry a procurement decision yet. Where no provider may process the data at all, the answer is still running a private model instead.

Plugins gain extensions: sidebar homes, interactive panels and file viewers, along with better tools for creating and finding plugins. Sites can now host plugins on Business, Enterprise, Healthcare and Edu, and MCP Events support lets plugins trigger automations.

MCP Events changes how a plugin starts work. Until now a plugin mostly answered when asked; with events, something happening in another system can start the work. That is how an assistant becomes part of a process, and it is also how an automation ends up running at 3am with nobody watching.

Plugins hosted on Sites give IT one place to publish internal tools to staff inside ChatGPT, which beats every team wiring up its own connector. It also makes plugin review a job someone has to own: what each plugin can read, what it can change and who approved it.

Business and Enterprise customers can now let people call @ChatGPT in Slack and Microsoft Teams channels without each of them holding a ChatGPT licence.

This changes the licence arithmetic. Staff who would use ChatGPT a few times a week are hard to justify a seat for, and now they can reach it where they already talk. It also means ChatGPT reads channel content in Slack and Teams, which is a data question as much as a productivity one.

Anthropic's nearest equivalent, Claude Tag, is a public beta for Team and Enterprise that works in Slack, runs each task in a short-lived sandbox and bills against an organisation usage balance with a spend limit an Owner sets. Its overview page describes Slack only, so a Teams-first organisation should check before assuming the two match.

Sign in with ChatGPT lets people use their ChatGPT plan allowance in 16 partner tools, including Cognition's Devin, Notion, Vercel, T3, OpenClaw and Dactyl. Pro 500 is a new plan with 25 times the Plus allowance, and it includes Ultrafast.

The OpenAI Marketplace lets eligible enterprises apply part of their OpenAI commitment to approved partner software. The first 32 partners include Figma, Adobe, Sierra, Decagon, HubSpot, Salesforce, ServiceNow, Harvey, Legora, Palo Alto Networks, CrowdStrike and Baseten.

The Marketplace is a procurement story. If you already hold an OpenAI commitment, software from those partners can draw it down, which can make a purchase simpler to approve. Read the commitment terms with that in mind: money spent on partner software is money that is not there for model usage later.

Sign in with ChatGPT moves identity the other way. It is convenient, but it gives company work a new route into third-party tools under a ChatGPT allowance rather than a contract your procurement team signed, which is what single sign-on and supplier reviews exist to catch.