OpenAI DevDay 2026: Everything Announced, What’s Available and What’s Coming Next

OpenAI wants a bigger place in your day. That was the message running through DevDay 2026, from personal AI agents that keep working while you’re away to shared documents, meeting notes and a subscription you can take into other apps.

The headline model was GPT-6.1 Sol. The announcement I’d pay more attention to was dots.

Dots put OpenAI’s ambition into plain sight: give an assistant an ongoing responsibility and let it get on with the work. Around that idea, OpenAI is building somewhere to keep the results, ways to involve other people and tools that can start a task without waiting for another prompt.

That is a much bigger bet than a better chatbot. It also explains why an event called DevDay had so much to say about how the rest of us might use ChatGPT.

There was plenty to get through on 29 September. The official recap has 25 sections, although calling them 25 brand-new launches would stretch the truth. Some features have arrived, some are rolling out, and others are still previews or promises. Here’s how they fit together, and which ones you can actually use.

Availability checked on 3 October 2026. Launch announcements and current help pages are distinguished below; account access may vary.

Key points

• Dots are the biggest change in how you use ChatGPT. They can take on an ongoing job and keep working between conversations.
• GPT-6.1 Sol makes capable AI cheaper to build with. It launched in Work, Codex and the API, with ordinary Chat still outside the launch.
• Space and Pages give the work a home. You can keep improving a document with other people and AI, rather than losing the result in a long chat.
• Your subscription can stretch beyond ChatGPT. Participating apps can use eligible Plus and Pro allowance, with your permission and limits.
• The launch list runs ahead of availability. Several features are rolling out or in preview; Sol Ultrafast was announced for later; collaborative slides and Private Inference are also future plans in the checked sources.

The main features at a glance

This is an access snapshot, not a promise that every eligible account already has every feature. Region, rollout timing and workspace settings still matter. The sections below explain the exceptions.

Dots are the announcement to watch

A dot is a personal AI agent powered by Astra, with its own cloud computer and access to the apps you choose to connect. It remembers the context of ongoing work and can keep making progress between conversations.

The appeal is easy to understand. Imagine you’re researching a big purchase. You could ask an assistant to compare the options once. Or you could give a dot responsibility for the research, ask it to keep track of changes and have it bring back the developments that might affect your decision. That is the kind of ongoing assignment dots are designed for.

For me, the hard part of that proposition is deciding what deserves attention. An agent that finds another 40 things for you to read has created homework. One that notices a meaningful change, explains it clearly and prepares the next step could be useful.

OpenAI has given dots controls for that relationship. Custom Rules let you define actions they can take, actions that need approval and actions they must avoid. Proactive background research uses read-only tools. Giving a dot access to your local computer is a separate choice and starts switched off.

Those details matter because an assistant with an ongoing job will encounter decisions you never thought to spell out. “Keep this project moving” sounds wonderfully simple until the project needs money spent, an email sent or a commitment made.

The first dot is included on eligible Pro and Business Premium plans, with an allowance for deeper work. Access is gradual. Pro availability currently excludes the European Economic Area, Switzerland and the UK; Business Premium covers supported ChatGPT regions. Enterprise, Edu and Healthcare users have an admin-enabled beta, initially off. The dot setup guide has the current eligibility details.

There is a separate enterprise idea here too: specialist dots with defined responsibilities inside an organisation. Those are starting in focused pilots. The planned Microsoft Agent 365 integration is still being developed.

I’d put dots at the top of this event because they change what OpenAI is asking us to trust an assistant with. We already know AI can produce a convincing answer. Keeping an ongoing responsibility on track is a more demanding job.

Sol brings the price down. Ultrafast brings the wait down

GPT-6.1 Sol is the new model, with OpenAI emphasising improvements in coding, computer use and professional work. The company puts its performance close to Astra on selected evaluations, at much lower standard prices.

That combination matters more than another impressive chart. If a less expensive model can do work that previously needed the premium one, developers have room to build products people can afford to use regularly.

The standard API rates are US$2 per million input tokens and US$10 per million output tokens for prompts with up to 272,000 input tokens. Tokens are the pieces of text a model reads and produces. Those rates are one-fifth of Astra’s standard input and output prices; other processing options have their own charges. The Sol announcement sets out the launch pricing and performance claims.

Sol launched in ChatGPT Work, Codex and the API. It was explicitly not in ordinary Chat at launch. The Work and Codex rollout starts with Pro, with Plus, Business, Enterprise and Edu following. That distinction is worth remembering if you open the usual chat window and wonder where the new model went.

Astra Ultrafast tackles a different irritation: waiting for a capable model to finish writing. OpenAI advertised up to eight times faster token generation in Codex and up to six times in the API at DevDay. A task that also spends time searching or using software will have a smaller overall improvement.

Ultrafast is a premium tier. Astra Ultrafast is available through the API and in Work and Codex on Pro 500 and Enterprise. Its API rates are substantially higher than standard Astra. Sol Ultrafast was announced as coming soon on 29 September.

I like the distinction between these two announcements. Cheaper intelligence makes more ideas viable. Faster intelligence makes the demanding ones less tedious to work with. They solve different problems, and OpenAI is charging accordingly.

Codex can keep working after you close the laptop

Codex Cloud is a straightforward improvement with an obvious benefit: a coding task can continue in a cloud environment while your computer is asleep. You can come back to the changed files and results, or check in from another device.

Reusable environments also mean a project’s tools and setup can be prepared once and used for subsequent tasks. Developers should spend less time explaining the surroundings before asking for the actual work. Cloud access covers Plus, Pro, Business, Healthcare, Edu and Enterprise, subject to the account’s repository access and workspace settings.

The refreshed Codex CLI, the version developers use in a terminal, makes longer sessions easier to follow. It adds a clearer view of multiple agents and improves prompt editing, picking up old sessions and handling separate pieces of work. Voice is part of that experience, although CLI voice was already enabled by default before DevDay.

The other half of doing more coding is having more code to check. Code Review brings summaries, comments and file changes together so people can inspect the work and question Codex about possible problems. Automatic reviews can make a first pass in the cloud. GitHub support is generally available; GitLab support remains in preview.

Codex Security Cloud goes further into security, scanning connected GitHub repositories, investigating findings and preparing proposed fixes. It can run on demand or keep checking as the code changes. Access is on Pro, Business, Enterprise and Edu, with the required setup and billing. Daybreak Blue model access here belongs to Security Cloud.

The useful pattern is that work can continue while the person steps away, then return in a form they can examine. That is a better proposition than simply generating more code and leaving someone to sort through it. Of course, a proposed fix still has to deserve its place in the project.

ChatGPT finally gives the work somewhere to live

Ask ChatGPT to help with a project over several weeks and the conversation can become its own filing problem. The answer you need might be halfway through a thread, with a newer version in another one.

ChatGPT Space and Pages address that mess. Space becomes a home for the project’s documents and files, while Pages are editable documents where you can write, research, add images and charts, and work with other people and agents.

The difference is practical. A plan can live in one document that everyone improves, rather than being rewritten in a fresh answer every time someone asks a question. Comments and shared editing give the result a life beyond the conversation that created it.

This is where the dots idea starts to make more sense. An assistant with an ongoing assignment needs somewhere to leave useful work. A Page can be the thing you return to: the current brief, the updated research or the plan with the latest decisions in it.

Space replaces Library for eligible Pro, Business and Enterprise accounts. Creation and editing are on web and desktop. On mobile you can find, read and share Pages, with editing still coming. Enterprise sharing is an opt-in preview. Sharing a Page gives people access to its contents, rather than opening up your private chat history.

The next piece is collaborative slides, where teammates and agents will be able to work on the same deck, leave comments and export it to PowerPoint or Google Slides. That experience is still coming soon, with Pro, Business and Enterprise availability planned in the coming weeks. The Space overview separates what is available from what is on the way.

I think this is one of the quieter announcements with more potential than its name suggests. Good AI output often needs editing, discussion and another pass. Giving that work a stable home could be more valuable than making the first answer a little prettier.

The assistant joins the conversation

OpenAI is also making the work less dependent on one person’s private chat.

Teams and Team Tasks let Business and Enterprise colleagues share work and maintain the instructions for recurring tasks together. A weekly project update, for example, can belong to the team, with people able to change what it should cover. Tasks can run on a schedule or respond to supported events in connected apps. Shared membership does not hand over everyone’s private accounts or memories.

@ChatGPT in Slack and Microsoft Teams moves the assistant into the conversations people are already having. Mention it in an enabled channel or thread, give it the job, then let colleagues add context and refine the result there. Business, Enterprise and Edu access requires administrator setup; colleagues in enabled channels do not all need individual ChatGPT licences.

That could remove a familiar bit of busywork: one person copying a discussion into ChatGPT, explaining it again and pasting the answer back for everyone else. The assistant can work with the group where the discussion happened.

The Meetings plugin connects spoken discussion to that same flow. It produces personalised notes and action items in Space, which can then help with a follow-up or an updated project plan. Notes start private, and participants need to consent before capture. Audio is deleted after the notes finish processing. Audio can remain temporarily on the Mac if upload or processing has not finished.

Meetings is currently a macOS beta for Pro and Business. Enterprise is in a limited alpha, while Windows, iOS and Android support are coming later. This is one of the features where the platform matters as much as the subscription.

Finally, shareable profiles provide a place to showcase creations such as Sites and plugins. They could make it easier to discover something useful a colleague has built. Enterprise profile sharing is still coming, the Sites showcase is on web and desktop, and personal profiles cannot be used to share Skills publicly.

Plugins start looking more like apps

A plugin becomes much more useful when you can actually work inside it. Plugin extensions give developers room for sidebar apps, interactive panels and viewers for the files their products use.

Picture asking ChatGPT to help with a design, then being able to adjust the result in a panel beside the conversation. That is the kind of experience these extensions open up. Access varies by surface, with Free and Go web extensions still coming soon.

OpenAI is also improving plugin creation, submission and discovery. Plugin Creator helps developers build, the submission process gives clearer feedback, and the directory and conversation recommendations aim to help people find relevant tools. Easier discovery will matter if good plugins are going to reach anyone beyond the people who already know their names.

Sites with plugins take the same idea into small apps you build with ChatGPT. Colleagues can use a shared Site with their own connected information and permissions. The DevDay recap lists Business, Enterprise, Healthcare and Edu for Sites plugins. The current Sites guide describes read-only connected-tool access in eligible Business and Enterprise workspaces, subject to permissions. Sharing an app does not mean sharing unrestricted access to its creator’s accounts.

Then there are MCP events, a technical name for a simple improvement: a connected app can tell ChatGPT that something has happened. A new project task might trigger a draft plan using the linked documents, while you’re doing something else. OpenAI is implementing webhook support from the proposed events specification; a connected service has to support the relevant event.

For most readers, the name of the protocol is beside the point. The change is that the assistant can respond when the work arrives. Together with dots and Team Tasks, it makes the background-work idea much more concrete.

The developer announcements, in plain English

There were three more ways for developers to put OpenAI’s capabilities into their own products.

The Decisions API uses Luna to choose between answers a developer defines in advance. It could classify an incoming request or decide which step an agent should take next, using text or images as context. Sometimes a product needs a quick choice between known options rather than a long explanation. This is in limited preview, with a broader release planned.

The Agents API computer-use update lets agents work through software in an OpenAI-hosted browser. They can navigate a website and interact with its interface, which helps with jobs such as testing a page or gathering information from an application. The Agents API itself entered public beta on 10 September; computer use was the 29 September addition.

Bedrock Managed Agents, powered by OpenAI, is the AWS version of that broader agent proposition. It brings OpenAI models and agent capabilities into AWS infrastructure, working with AWS resources and permissions. It is an AWS preview, with its own access and commercial arrangements, rather than a feature included in a personal ChatGPT plan.

These announcements give developers different pieces of the same puzzle: making a choice, carrying out a sequence of actions and running the resulting product in the environment their organisation uses. Readers do not need to learn every API name to understand why those pieces matter.

Privacy gets a more specific promise

Private Intelligence covers OpenAI’s work on keeping sensitive information protected while using its models.

The available part is Zero Data Retention with Private Safety Processing. It enables automated safety review in a protected environment designed to keep the underlying content inaccessible to OpenAI personnel. Selected encrypted safety records are retained in customer-controlled storage for 30 days. “Private” needs that explanation; it should never be casually translated into “nothing is stored anywhere”.

The future part is Private Inference, which uses confidential computing and verifiable controls. OpenAI announced a preview for the northern autumn. It has not arrived as a generally available feature for every customer.

This will be a more consequential announcement for organisations handling sensitive material than for someone asking ChatGPT for a dinner idea. The value lies in the specific arrangement and who can access the information, rather than the reassurance of the product name.

Your subscription can travel, and Pro goes up to US$500

The more consumer-friendly access announcement is using your ChatGPT plan in other apps. Eligible Plus and Pro users can allow participating products to draw on the AI usage already included in their subscription.

Signing in with a ChatGPT account predates DevDay. The new development is plan usage across participating partners, including products such as Devin, Notion and Vercel. That is in limited preview. You choose whether to allow it and can set how much of your weekly allowance each app may use.

It is an appealing idea if you already pay for ChatGPT and want to use that allowance somewhere else. It also gives you a reason to watch consumption: usage in a partner app draws from the same pool. The partner can still charge for its own product, and purchased credits require a separate opt-in. The plan-usage guide explains those choices.

At the expensive end, Pro 500 adds a higher allowance and Astra Ultrafast. Here’s how the three Pro tiers compare, using OpenAI’s Pro plan details.

Pro 500 has the largest included allowance. Eligible existing Pro 200 subscribers keep their previous allowance through 29 October 2026, then move to a lower allowance at the same US$200 monthly price. Buying extra credits on Pro 100 or Pro 200 does not unlock Ultrafast. Use the checkout amount for your own purchase rather than reading these as Australian-dollar prices.

US$500 a month is a serious subscription. For someone doing demanding work throughout the day, the combination of more usage and less waiting could make sense. For an occasional ChatGPT user, nothing in this announcement makes it the obvious upgrade.

OpenAI Marketplace is a different commercial programme altogether. Eligible enterprise customers can put part of an existing OpenAI spending commitment towards approved partner software. It is about buying business products under specific agreements, rather than a consumer app store or an extra perk on Pro.

What is still coming?

Some of the most appealing announcements are still ahead of us. Here’s the short version of the promises and early access, so you can tell what belongs on a wish list and what you can investigate now.

The big question is how much attention this saves

Taken together, DevDay’s announcements describe a much larger role for ChatGPT. A dot keeps a project moving. A connected app tells it something has changed. Codex does the coding work. Pages hold the results. Other people join in where the discussion is happening.

That is a coherent direction. It is also a lot of product to roll out at once, with plenty of gaps between the stage announcement and the version an individual account can use today. Sol Ultrafast was announced for later; collaborative slides and Private Inference are also future plans in the checked sources, and several of the features that have arrived are still in previews or gradual rollouts.

The standard I’d judge all of this by is fairly simple: how often do you come back to useful progress, and how often do you come back to something that needs rescuing?

An assistant can produce plenty of activity without making your day easier. The interesting promise here is work you can leave alone for a while and feel pleased to return to. Dots will carry a lot of that expectation. If OpenAI gets the everyday experience right, that will be the part of DevDay people remember long after the model numbers blur together.