OpenAI’s DevDay produced more than 20 announcements. The three I’d pay attention to are Dots, its ongoing AI agents; Space, a place for people and agents to work together; and GPT-6.1 Sol, a model that has been getting serious work done without chewing through much of my plan allowance.
I’ve been using all three. Together, they show what OpenAI is trying to build: a place where the context you’ve accumulated with AI keeps producing useful work, and where colleagues can join in.
An agent becomes more useful when it understands the work you’re doing. A shared workspace lets other people contribute what they know. A cheaper model lets you repeat more of that work without exhausting your budget. OpenAI is trying to make all three happen in the same place, until bringing the next assignment into ChatGPT feels natural.
[LINK: Nate’s Library MCP]
Your subscription now works inside your AI.
Nate’s Library MCP is a read-only MCP server that gives your AI direct access to my entire published content library. Connect once, and then the archive just shows up in your normal AI conversations. You ask your AI a question during your actual working session and it pulls from everything I’ve published to answer it.
Setup takes just a few clicks.
[LINK: Join the Slack →]
It’s where I’ll be sharing things between articles, where you can get help on builds in real time, and where the fastest conversations in this community are already happening. I’ll see you in there!
Both are included with your paid membership. If you have any issues connecting to either platform using the links provided, please email support@natebjones.com.
What ChatGPT Dots can take over first
My Dot caught a recurring meeting that kept landing on weekends and dealt with it. I’d been too annoyed to clean it up myself. And who likes weekend meetings? It also spotted a conflict between a DMV appointment and my travel plans.
Those are small jobs. They’re also an early example of the thing I care about: AI using what it knows to notice work I haven’t explicitly asked it to do.
A Dot runs on GPT-6 Astra, has its own cloud computer, and can work through connected applications. Its memory includes ChatGPT memories, recent conversation context, and information from connected apps. It doesn’t automatically inherit every previous chat. You choose what to connect and what responsibilities to give it. I like that boundary.
Of course, Meta’s Muse also has a computer, memory, and a free entry point alongside paid options. That makes “does it have a computer and memory?” a poor buying question. I want to know which responsibility an agent can take over, what it needs to know, and how much work I still have to do to get the result.
With Dots, I’m particularly interested in responsibilities built around substantial AI context. You’ve explained a project, corrected assumptions, and worked through what a customer actually needs. The longer play is to make that understanding useful between conversations, so you don’t have to reconstruct the assignment every time something changes.
OpenAI’s launch example makes this real. A product’s scope changes, and the Dot prepares revisions to the launch materials. In practice, a delayed feature can also affect the announcement, a sales deck, and what somebody promised a customer. Knowing which promises depended on that feature is where the useful work begins.
The payoff is a team arriving at its next customer conversation with an accurate proposal and a clear account of what changed. Somebody still decides what to promise. The agent can do more of the work needed to make that decision well.
Give your Dot a recurring responsibility in an area where the relevant information is available. Watch whether it becomes useful enough that you keep bringing the work back.
Space gives the team somewhere to continue
The launch example also explains why Space and Pages belong in the same announcement. Colleagues and AI can work on shared documents. The salesperson can flag a customer commitment while the product lead explains what changed. The next revision can use both.
A lot of AI use still depends on one person doing the explaining. That person knows the history, knows which answer is wrong, and knows the exception that makes the obvious solution inappropriate. If the work then moves to a colleague, much of the explaining starts again.
This is the practical continuation of the AI-native workplace question. How does one person’s reasoning help the next person doing the job? Shared pages give that reasoning somewhere to stay, but the value depends on people making the relevant knowledge available, including why a decision was made.
Space and Pages are available on Pro, Business, and Enterprise. Collaborative slides and spreadsheets are coming. Business and Enterprise also get team tasks for recurring work and ChatGPT participation in Slack and Teams. Sharing a Page does not automatically share your private chats or memories, though anything written on the Page, including details drawn from memory, is visible to collaborators.
This puts OpenAI into the working day Microsoft is pursuing with Copilot and Autopilot, and Anthropic with Claude Tag. Microsoft starts with employers already buying its working environment. Claude Tag works with teams in Slack. OpenAI is giving people who already work extensively in ChatGPT a reason to bring colleagues into it.
My starting point would be where the team’s decisions, documents, and people already are. Moving work needs to buy you something. If ChatGPT holds substantial project knowledge and Space lets the team use it, that is a concrete reason to try it. If the necessary information and permissions are elsewhere, the agent needs a workable route to them.
Try GPT-6.1 Sol before buying more capacity
Once you give AI recurring work, cost becomes harder to ignore. An occasional impressive answer and a job you want done every day make different demands on your plan.
Sol has been one of the most useful parts of this release for me. I’ve used it for documents, code, and research, and it has felt close to Astra while consuming very little of my plan allowance.
For heavy users on paid plans, that is often the cost question: how much work can I get done before I hit my limit?
OpenAI reports that GPT-6.1 Sol approaches Astra on selected evaluations at one-fifth of Astra’s standard input and output token prices. It is available in Work, Codex, and the API, not yet in ordinary Chat. My experience is encouraging. It isn’t a controlled benchmark. As with Opus 5.5, I care about the work a model finishes for the cost.
I still want Astra for the most demanding assignments. But if Sol can handle a recurring job well, using it leaves more of my allowance for work that needs the more expensive model. Model choice becomes part of managing how much I can get done.
The Pro tiers are $100, $200, and $500 a month. Each includes a first Dot where eligible. The higher tiers buy more usage, with the largest allowance and Astra Ultrafast on $500. You don’t need that top plan just to get a Dot. Pro’s Dot rollout currently excludes the UK, Switzerland, and the European Economic Area.
Existing $200 subscribers should check their allowance change: eligible subscribers keep the previous allowance through October 29, then move to a lower allowance at the same price. New subscribers outside that temporary arrangement already get the lower allowance.
Ultrafast has its own tradeoff. It offers up to eight times faster token generation in Codex, which does not mean every job finishes eight times faster. It also consumes included allowance at eight times the standard rate.
Would I pay the extra $300? First I’d try Sol on more of my existing work. Then I’d look at what the remaining limits or waiting time actually prevent me from doing. If an hour of usable working time is worth $100 to you, three recovered hours a month cover the difference. A job finishing at 3 a.m. instead of 4 a.m. may recover no usable time at all.
What a smaller team could build
The same economics matter if you’re building something for customers. Work that is too expensive to perform frequently can become a viable service when capable models make it affordable to repeat.
Codex Cloud provides reusable environments for building and maintaining software. The Agents API adds computer use through a hosted browser. Plugin extensions let a developer put interactive panels inside ChatGPT, where a customer can inspect and work with the results.
Return to the sales team from the launch example. A company may already pay for AI and still lack the time or expertise to keep product changes, sales materials, and customer commitments aligned. You could build a service that notices a change, finds the affected material, and prepares revisions for the people who can approve them.
You would need to establish which changes the service handles reliably and if the time saved justifies the price. A proposed change to a contractual promise should stop with someone authorized to make that decision.
Someone else can copy the panel you build. They still have to understand the customer’s exceptions, know what a good result looks like, and earn enough trust to become part of the work. That is where experience in an industry can turn into something useful to sell.
This is why I think these releases should make small teams more ambitious.
The recurring job somebody trusts you to get right is what makes the product worth keeping. Form factor is converging. Utility is not.
Coming up
If more of your context ends up living in ChatGPT, a harder question follows: what happens when you want to leave? Tomorrow’s Executive Briefing looks at how better agents are changing what software is worth. When an agent takes over a job, the software you used for it gets easier to drop. The agent you’ve spent months training gets harder to replace. I’ll show you how to settle that with evidence, and if you do have to switch, how to make it a lot less painful.
