OpenAI DevDay 2026: who will be working for whom?
OpenAI has introduced personal agents called dots, a cheaper GPT‑6.1 Sol and more ways to use your subscription outside ChatGPT. It promises AI that takes on more work. Its success will also depend on how much supervision it needs in return.
At a glance
At DevDay on 29 September 2026, OpenAI introduced more than twenty updates. Personal agents called dots run on GPT‑6 Astra and have their own cloud computer, memory and connections to apps. Their initial availability on the personal Pro plan excludes the Czech Republic and the rest of the European Economic Area; Business Premium has broader regional availability. GPT‑6.1 Sol costs two dollars per million input tokens and ten dollars per million output tokens in the API. According to OpenAI, it approaches Astra in several areas. It is available in Codex, ChatGPT Work and the API, but not yet in regular Chat. Sign in with ChatGPT lets users draw on their Plus and Pro subscriptions in sixteen partner tools, including OpenClaw. The new Pro 500 costs 500 dollars a month. Pro 200 receives a lower usage allowance, while eligible existing subscribers can keep their original allowance until 29 October 2026. The commentary judges the announcements by agents’ ability to finish work and save human attention, rather than by the number of integrations.
Personal AI now has its own computer, memory, access to apps, and the ability to work when you are away from the keyboard. OpenAI calls it dots. The idea is appealing: you set a goal, the agent gets to work, and it comes back with a result or a decision that really needs you.
At DevDay on 29 September, OpenAI announced more than twenty updates. Alongside dots came GPT‑6.1 Sol, shared workspaces, Codex extensions, and a subscription costing 500 dollars a month. Particularly interesting is the option to use your paid ChatGPT allowance in partner tools. OpenClaw is among the sixteen named partners. [1]
Together, these announcements show what OpenAI is aiming for. It wants to be the place where you assign work, where it happens, and where you collect the result. More than a little window you copy an answer from.
Dots are meant to keep work going between conversations
Dots run on GPT‑6 Astra. They have their own cloud computer and browser, can use connected apps, and can learn from ongoing feedback. OpenAI says its plugin ecosystem connects them to more than four thousand apps. You should be able to talk to them in ChatGPT, Slack and Microsoft Teams. [2]
Continuity is what matters. The agent is meant to retain a project’s context, continue between conversations, and handle several tasks at once. With permission, it can also work on your computer.
Just watch the marketing. “Works 24/7” does not mean it will simply start sending proposals to clients at three in the morning. Proactively exploring connected apps is limited to reading. Actions have rules and approval requirements. Working independently on an assigned task and looking for opportunities in the background are two different things. [2]
And then there is the Czech reality: the initial release of dots on the personal Pro plan excludes the European Economic Area, the UK and Switzerland. That means the Czech Republic too. Business Premium is meant to be available in all supported ChatGPT regions; on Enterprise, access depends on the administrator enabling it. Even there, the feature may not appear for all accounts at once. [3]
Sol 6.1 may be a more practical update than the most expensive model
According to results published by OpenAI, GPT‑6.1 Sol approaches Astra in coding, computer use and professional work. Meanwhile, its standard input and output token prices are one fifth as much: two dollars per million input tokens and ten dollars per million output tokens. Reused cached input costs ten cents per million tokens. [4]
Tokens are the units a service uses to count the text it processes. For an agent that repeatedly reads context and takes several steps in succession, their price quickly makes a difference.
This is good news for anyone who wants to do things with AI regularly. A token costing one fifth as much does not, however, automatically mean a finished project costs one fifth as much. That depends on the number of attempts, the tools used and the corrections needed.
Sol 6.1 has been available since the announcement in the API, Codex and ChatGPT Work for Plus and higher supported plans. It is not yet available in regular Chat. [4]
ChatGPT is building its own office
ChatGPT Space and Pages add a shared space and live documents that people and agents can work on together. Text, charts, research and project materials no longer have to end up buried in the thirtieth conversation thread.
Space and Pages are starting to become available for Pro, Business and Enterprise. Presentations edited together, with export to PowerPoint or Google Slides, are due in the coming weeks. The Meetings plugin is entering beta on macOS for Pro and Business: it takes meeting notes and saves a summary with tasks. [1]
Developers are getting reusable cloud environments for Codex, voice input in the terminal, and a clearer view of agents working concurrently. There are new ways to review code changes, plus Codex Security Cloud, which scans repositories, investigates security findings and prepares fixes.
Plugins, meanwhile, can gain their own interactive panels and editors directly in ChatGPT. Automations will be able to run in response to events in connected apps. The Agents API is getting computer use and further capabilities for multiple agents; the partnership with Amazon brings managed agents to AWS. [1]
I consider this combination more important than the list of features itself. An agent needs a place where work can be found, taken over and finished. Otherwise, it produces another piece of text that somebody has to move elsewhere by hand.
The subscription opens up. The pricing gets tougher
Sign in with ChatGPT is meant to let you use your Plus or Pro allowance in sixteen partner tools, including OpenClaw, Devin, Notion and Vercel. Users can control how much each tool consumes. [1]
That is a useful opening of the subscription. You can choose a different working environment and use something you already pay for within it. Be careful: this does not mean an unlimited API for free. Supported use counts towards your allowance.
Follow the money. The new Pro 500 costs 500 dollars a month and includes Astra Ultrafast. At the announcement, this faster variant is also available in the API and in Work/Codex for Enterprise. Sol Ultrafast is due to follow later. [1, 4, 5]
The less pleasant news is tucked away under Pro 200. New subscriptions have a lower usage allowance. Eligible existing subscribers can keep their original allowance until 29 October 2026, then they too will move to the lower one. The price stays at 200 dollars a month. [5]
A more efficient model really may do more work while consuming less. But until I see that on my own tasks, a lower limit is still a lower limit. Cheaper model operation and a better-value subscription for the customer are not the same thing.
An agent must not need a full-time manager
The direction OpenAI has shown makes sense to me. A personal agent with memory, tools and long-term context is much more interesting than endlessly explaining who I am and what I dealt with last time. And I welcome the option to use my subscription in another tool.
But I would not measure dots’ biggest promise by the number of connected apps. What use are four thousand integrations if somebody has to check every ten minutes whether the work has got stuck?
Take an ordinary project handover. A good agent finds the current materials, separates them from old versions, identifies missing information and prepares an output somebody else can actually take over. A weaker agent sends an impressive plan, a few progress reports, and finally asks whether it should continue. Both can look capable in a demo. You find out the difference when you need to leave the computer.
I do not expect an agent never to consult me about anything. I want it to recognise which decisions are worth interrupting a person for, and to finish routine work within the agreed boundaries. Taking on a task also means taking on the responsibility for what happens to it.
Alongside reliability, there is the question of choice. The more personal context and working habits an agent carries, the more important the ability to change tools or providers will become for me. Signing in through ChatGPT deals with access and billing. On its own, it does not yet answer how I can take months of shared work with me.
OpenAI now has a product on which these promises can be tested. Let it show that, after a week with an agent, there is less work left on the desk and fewer unfinished tasks in my head. I do not need another app I have to keep an eye on.