At DevDay on 29 September, OpenAI showed not just another model but a whole working environment: always-on agents, team spaces, Codex in the cloud and new ways to connect apps. The common idea is fewer one-off conversations with an AI and more work that carries on between them.
Below are the eight main threads of the announcements: what each one is, why it’s needed, and where the product you can already use ends and the promise begins. Terms and availability are given as of 30 September 2026, based on OpenAI’s official materials; benchmark results and speed figures are the company’s claims, not our own independent measurements.
1. Dots: an agent you hand ongoing work to
Dots are always-available agents running on GPT-6 Astra. Each has its own cloud computer, a browser and access to the apps the user has connected. They’re pitched not only for one-off errands but for ongoing responsibility: monitoring feedback, preparing fixes, updating project materials as new data comes in.
In the Dots announcement OpenAI describes a developer scenario: the agent notices recurring customer requests, picks out small improvements, writes and tests the changes, and then brings finished PRs for review. For a content creator, a similar process starts with a new interview transcript and ends with draft posts and episode notes.
The difference from an ordinary chat is continuity. You don’t have to explain the project from scratch every time or stitch separate tasks together by hand. You can talk to the agent in ChatGPT, Slack and Microsoft Teams; the company says context is kept across channels.
But “works around the clock” doesn’t mean “can do anything without asking”. Proactive background research is limited to read-only tools. For actions there are access rules, automatic checks and approvals; sensitive steps, such as changing a password, stay with a human.
Who gets it: the rollout is gradual. On Pro, in all markets except the European Economic Area, Switzerland and the UK; on Business Premium, in every region where ChatGPT operates. In Enterprise, Edu and Healthcare the beta is off by default and is switched on by the workspace administrator. The first dot is included in an eligible plan, but the Codex and ChatGPT Work tasks it launches draw on your normal usage limits. This isn’t an unlimited employee hidden inside a subscription.
2. ChatGPT Space: a team space, not just a notebook
ChatGPT Space brings people, ChatGPT and Dots together around shared knowledge and materials. A team creates a dedicated project space, and ChatGPT helps keep it in order according to the instructions it’s given.
Alongside it comes a new entity, Pages: documents for people and agents to work on together. On a page you can write, do research, and create charts and images, and colleagues can add to the result and leave feedback.
The practical point is that you stop forwarding fragments of different conversations between team members. The project gets a shared place where the materials live and where work can pick up from where it stopped.
Who gets it: Space and Pages are announced for Pro, Business and Enterprise in the desktop app and on the web. On mobile, Space offers search, reading and sharing pages; creating and editing are still only promised. Shared presentations with simultaneous editing are also announced for the coming weeks, not handed to everyone on the day of the keynote.
This continues the push towards merging ChatGPT, Work and Codex into one app: now team collaboration is being added to the shared shell.
3. GPT-6.1 Sol: closer to Astra, cheaper in the API
GPT-6.1 Sol is an update to Sol for agentic coding, operating computer applications and professional tasks. OpenAI positions the model as close to GPT-6 Astra on a number of capabilities, but with standard API input and output token prices at one fifth of Astra’s.
The official model announcement gives prices per million tokens:
- input — $2;
- cached input — $0.10;
- output — $10.
That’s the standard rate for requests where the input context is no more than 272K tokens. If the input is longer, the whole request costs more: input and cache double, output goes up by half. Batch and Flex modes are half the standard price, fast mode is twice the price. Full terms are in the API documentation.
If you’re a developer, that’s a good reason to test Sol on your own tasks. If the agent reuses shared project context many times, the price of cached input matters too. But the price of a token isn’t the price of a finished result: retries, reasoning length and errors can change the final economics.
And don’t read “almost Astra” as “Astra is no longer needed”. OpenAI itself notes that Astra remains stronger on the hardest scientific tasks. Results on individual benchmarks don’t guarantee the same quality on any given repository or working document.
Where it’s available: in the API as gpt-6.1-sol, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. An important caveat from the detailed announcement: in the regular Chat mode the model isn’t available yet. Having a subscription and seeing the model in your usual chat list are different things.
4. Codex in the cloud: work no longer tied to an open laptop
Codex can be run locally, remotely from a phone, or in the cloud from different devices. For teams there are now reusable development environments with consistent settings and permissions.
The point isn’t just being able to type a request on your smartphone. When the environment is reproducible, the agent doesn’t have to rebuild it every time, and colleagues don’t have to work out which dependencies and permissions were used in the previous run.
Who gets it: Plus, Pro, Business, Healthcare, Education and Enterprise, according to the official DevDay recap.
Other parts of Codex were updated around this. The CLI gained voice control of tasks and an /agents view for delegating and monitoring several jobs. The desktop app has a revamped Code Review that walks through the changes, shows the diff and lets you ask Codex questions before sending feedback to GitHub or GitLab. Automatic reviews can run in the cloud while the user is away.
Separately, OpenAI introduced Codex Security Cloud: scans of entire GitHub repositories on demand or on a schedule, analysis of new commits, triage of findings and preparation of fixes in the cloud. It’s available on Pro, Business, Enterprise and Edu in the desktop app and on the web. It’s a tool for defenders, not a reason to drop regular review and security checks.
Process still matters more than a flashy launch. Working with a cloud agent calls for the same engineering guardrails and definitions of done as a local one: a bounded task, tests and a verifiable result.
5. Decisions API: choose an action, don’t write a long answer
The Decisions API uses Luna’s capabilities for tasks with a finite set of predefined answers. The developer passes in context (text or images) and defines the allowed options.
For example, a request has to be routed to support, sales or the technical team. Or an agent has to choose its next step: ask for clarification, call a tool or hand the task to another agent. These are example uses, not ready-made integrations that OpenAI will set up instead of the developer.
What sets it apart from an ordinary generative request is the shape of the task. What you need isn’t free text but a constrained decision that a program then uses. That’s handy for classification, routing and choosing the next action in an agentic process.
Status: limited preview. The DevDay recap promises a broad opening in the coming days; for now you can’t assume the API is generally available to every account.
OpenAI also extended the Agents API with computer-use support. You can bring multi-agent workflows, search and tool calling, and context compaction into your app, with OpenAI running the infrastructure. It’s available through the API, and in Codex and ChatGPT Work on Pro 500 and Enterprise. This is a neighbouring announcement, not another name for the Decisions API.
6. Pro 500: higher limits, but not an endless quota
The new Pro 500 tier is designed for heavy use. In the DevDay recap OpenAI claims usage limits 25 times higher than Plus and access to Ultrafast. The price is $500 a month, as stated in OpenAI’s help article on Pro tiers; The Verge’s DevDay report says the same.
You can’t derive a specific number of requests or working hours from the multiplier: different models and tasks consume the included allowance differently. For heavy agentic workflows it’s more important to understand the real consumption of your own scenario than to go by the number in the plan’s name.
Status: OpenAI has announced the tier as available. Check the specific purchase terms, included usage and feature access in your own account: an announcement doesn’t override market restrictions or the phased rollout of products.
7. Ultrafast: paying extra for speed, not a new model
Ultrafast is a premium speed mode. In the official recap OpenAI claims generation of up to 300 tokens per second for GPT-6 Astra in Codex, up to an eightfold speed-up of generation there, and up to a sixfold speed-up in the API.
The key word is generation. This isn’t a promise to finish any task eight times faster: an agent may be waiting on a build, tests, a browser or an external service. Producing tokens faster and finishing a project faster are related, but not identical, measures.
Nor does Ultrafast automatically mean a higher-quality answer. Its job is to cut the wait where speed really matters.
Where it’s available: Astra Ultrafast is announced in the API, and in ChatGPT Work and Codex on Pro 500 and in eligible Enterprise and Edu workspaces. On Pro 500 the mode draws on included limits and credits; in Enterprise and Edu, on workspace credits, and access depends on the permissions granted by the administrator. On Plus, Pro 100, Pro 200 and Business it isn’t available at launch, not even for credits. Sol Ultrafast is promised soon. Don’t confuse the Astra mode that has already launched with the future Sol mode.
8. Plugins and the Marketplace: two different stories
Plugin extensions let you build your own panels in the ChatGPT sidebar, interactive interfaces next to the conversation, and viewers for files the product supports. It’s no longer just about connecting a tool so the model can call an API, but about giving the user a full working surface inside ChatGPT.
OpenAI has also made it easier to build and submit plugins, and updated how they’re discovered and recommended. The user chooses which plugins to connect and approves each one’s access. The main updates to extensions and the catalogue are announced for all plans.
Two more related changes: supported plugins can be added to Sites, and MCP events let you trigger automation when something changes in a connected app. For example, a new task on a project board can prompt a plan based on the related documents. OpenAI specifically calls MCP Events a proposed specification, not a finished standard.
OpenAI Marketplace isn’t the same as the plugin catalogue. It’s a corporate programme under which eligible enterprise customers can direct part of their already committed OpenAI spend to approved partner software. The first partners named include Adobe, Figma, Salesforce, ServiceNow and other vendors. For a corporate buyer this is a story about procurement and budget, not about finding a new extension in the chat.
What else was shown and what to do with it
The official recap also included other announcements: ChatGPT in Slack and Teams, team tasks on schedules and events, a meetings plugin, signing in to third-party tools via ChatGPT, and Private Intelligence for companies. So the eight threads above are an overview of the main points, not a complete list of everything new.
Taken together, the picture is clear: OpenAI is moving ChatGPT from a window for questions to an environment where people and agents run projects and developers build their own tools and interfaces.
Don’t start by handing an agent the whole company; start with one bounded process. For a developer: compare GPT-6.1 Sol with your current model on real tasks and try Codex in the cloud. For a team: try a shared space for one project. For Dots: pick recurring work with clear permissions, approval points and success criteria.
And check availability separately. “Shown at DevDay”, “opened as a preview” and “already on my plan” are three different states of a product.
