My interpretation of these announcements: OpenAI is trying to jump into the fast-growing bot market that is forming right now, where Meta’s Muse and Grok Bot recently took the initiative. And it keeps pressuring Anthropic in developer tools.
Dots take over ongoing tasks
Dots are agents powered by GPT-6 Astra, each with its own cloud computer and browser. They are meant to work continuously, use connected apps, and gradually learn the user’s preferences. Communication happens in ChatGPT, Slack and Teams, including voice calls.
OpenAI gives a developer scenario: Dots monitors customer feedback, prepares small fixes, tests them, and submits pull requests with a video of the changes. Not bad, but it will depend on the execution, meaning what “monitors customer feedback” actually amounts to. The service is not available in the EU, the UK and other European countries, so you have to connect through a VPN. It works nicely on the web; I don’t have it in the mobile or desktop app yet, but it’s supposed to arrive…
The boundary of autonomy is important: proactive background research uses read-only tools. The ability to change data or send messages is subject to rules and action review. Users can set their own rules and monitor activity. Dots gets access to the user’s own computer only if the user actively connects it.
Dots are rolling out gradually for Pro and Business Premium in selected countries; enterprise and education accounts can use a beta enabled by an admin. Specialized company dots with their own identity are still in pilots. Introducing Dots.
Compared to Meta’s Muse, there is a noticeable focus on the work process. While Muse gives you a personal assistant, Dots is aimed more at helping with work. But take this as a first impression; OpenAI will surely try to find other models of working with Dots. It strikes me that the model is needlessly powerful for everyday work, but I can’t say right now what that says about the product. Maybe about the haste with which the service was presented.
And just as an aside, on how billionaires throw their money around: SpaceXAI bought the domain Dot.com the day before the event and pointed it at Grok Bot. The price tag? Almost half a billion Czech crowns (20 million dollars). The annual budget of Brandýs (a town near Prague)…
Codex expands cloud development and change review
Codex Cloud lets you assign development tasks in a cloud environment containing the necessary repositories, tools and permissions. Each task has its own workspace and can keep running even while the user’s computer is asleep. Results can be reviewed and further edited from the web, mobile or desktop app. In other words, something we already know from Anthropic; no revolution here. Codex Cloud documentation.
The refreshed Codex CLI adds voice input and task control, a /agents overview for tracking delegated work, and changes to session resumption, prompts and working with worktrees. DevDay recap. The CLI still lets you browse and edit local code, run commands, and plug Codex into scripts and CI. Some of this was already available before, so among other things this is a cleanup and systematization. CLI documentation.
The new Code Review environment in the desktop app brings the review of finished code into one place. Version control support: GitHub is generally available; support for GitLab merge requests has preview status in the documentation. Code Review documentation.
Codex Security Cloud interested me the most. It adds security review of connected GitHub repositories and monitoring of new commits. Scans run in the cloud; for local repositories there is a separate Codex Security plugin. I just ran it locally on Vibecoding.cz and I’m curious what it will say. So far it’s not quite clear to me why it runs on the GPT-5.5 model; probably it stayed in my settings somewhere. The version in the desktop app correctly ran on 6.1-Sol, as I would expect. Security Cloud setup.
Support for working with PRs and code review strikes me lately as a very important direction, and I’m glad Codex is stepping up.
Agents API opens up infrastructure for custom agents
Agents API was already introduced on September 10, 2026 as a public beta. It gives developers access to Codex’s control layer: management of long sessions, context handling, tool use, and coordination of sub-agents. OpenAI provides the infrastructure in which agents can work with files, run code, and store intermediate results. Original Agents API announcement. But it’s more for enterprise development, in case you want to tie yourself closely to OpenAI’s infrastructure; it’s not the kind of free-spirited development we like here in Czechia…
The old-new DevDay news is support for computer control. Pay attention to it, it’s useful, and OpenAI has extended it to custom agents. Computer use documentation.
GPT-6.1 Sol changes the cost of longer workflows
The new GPT-6.1 Sol focuses on agentic programming, computer control and professional work. OpenAI claims performance close to GPT-6 Astra at one fifth of the standard input and output token prices. However, this is a vendor comparison based on its own evaluation, not proof of equal performance across all applications. Let’s wait for real-world experience.
According to the announcement, repeatedly used input stored in the cache costs $0.10 per million tokens. For agents this item is relevant, because across individual steps they may return to the same instructions and context. Introducing GPT-6.1 Sol.
A lower token price can allow more iterations and checks within a given budget. But the cost of a completed task also depends on the number of steps, tool usage, and the fixes required. The one-fifth price figure therefore can’t be automatically converted into one fifth of the cost of developing an application. I’m cautious here, because in practice it simply works out that way far too rarely…
Plugins get their own interface and event reactions
Plugin extensions let developers add apps to the ChatGPT sidebar, open them next to a conversation, and build viewers or editors for supported files. The extensions also include plugin settings, deep links, context sharing between the model and the app, and structured forms. The documentation notes that web extensions for Free and Go are yet to come; mentions in the message editor are available only in the desktop app. Extensions documentation.
For your own product, in my reading, this means the ability to connect a conversation with a purpose-built interface: the user can describe a request to the agent and then edit the result in the app’s panel.
MCP Events add automations triggered by an event from an external system. The user specifies what to watch and how to react; the server then sends the corresponding events to a callback. The documentation gives examples such as processing bug reports or document comments.
The support is based on a draft of the MCP Events specification. The integration uses webhooks and callback verification; it doesn’t support all the delivery methods listed in the draft. MCP Events documentation.
For agents that work continuously, the trigger is as important as the ability to write code: it determines when they should start and which request they should handle.
Other DevDay announcements
There are a few other things worth a brief mention. First of all, you can use Sign in with ChatGPT. And you can use your ChatGPT limits and subscription in that service, although by itself it doesn’t transfer money or your subscription to the service. Limits depend on the particular service and its configuration: it may be a number of requests, available models, or API usage. Any fees are charged according to that service’s terms; signing in does not mean automatic consent to pay.
ChatGPT Spaces are topic-focused workspaces in ChatGPT where you can organize conversations, files and instructions for a specific project or purpose. For example, I’m writing this article in Spaces. Somewhere I read that it will kill Google Docs. Let’s be clear: it’s a long way from that. It’s more of an AI collaborator than a text editor, so a different purpose, a different focus.
The remaining announcements:
| Announcement | Announced feature |
|---|---|
| Ultrafast | Higher token generation speed |
| Private Intelligence | Data protection including Private Safety Processing and planned Private Inference |
| Decisions API | Choosing among predefined answers; limited preview (a response to Jev) |
| Bedrock Managed Agents | OpenAI agents running on AWS |
| Plugin Creator and catalog | Easier creation, submission and discovery of plugins |
| Plugins in Sites | Apps with connected data and per-user permissions |
| ChatGPT Space and Pages | Shared workspace and documents for people and agents |
| Collaborative presentations | Simultaneous editing by people and agents; announced for the coming weeks |
| Teams and team tasks | Sharing work and recurring automations |
| ChatGPT in Slack and Teams | AI directly involved in team conversations |
| Meetings | Meeting notes and tasks |
| Shared profiles | Presentation of Sites and plugins |
| Sign in with ChatGPT | Sign-in and use of limits in connected tools |
| Pro 500 | New plan with higher limits |
| OpenAI Marketplace | Applying part of enterprise commitments to partner software |
Full recap and availability of individual announcements.
With Sign in with ChatGPT, it’s worth distinguishing between signing in and drawing on your subscription. The user can use the service purely as a sign-in identity. If they allow use of their plan, eligible requests from the partner app count toward their limits for ChatGPT Work and Codex. It is not a separate additional usage allowance, nor does it hand over an API key. The partner’s fees may be billed separately. Terms for using your plan in other apps.
According to the US price list, the Pro 500 plan costs $500 per month and includes Ultrafast. The documentation also describes changes to Pro 200 usage allowances for new and some existing accounts, so price alone isn’t enough to compare the plans. Overview of Pro plans.