OpenAI DevDay: Dots, Space and GPT-6.1 Sol, toward a ChatGPT that retains context and is shared across a team

By Nate (Nate's Substack).

OpenAI's DevDay produced more than 20 announcements, but Nate focuses on the three he considers most relevant: Dots (continuous AI agents), Space (a place for people and agents to work together) and GPT-6.1 Sol, a model that, in his experience using it, performs solidly while barely consuming the quota on his paid plan. The author explains that he has tried all three and that, together, they reveal OpenAI's underlying bet: to build an environment where the context accumulated with AI keeps generating useful work over time, and where teammates can join in on that work.

On ChatGPT Dots, Nate offers his own modest examples: his Dot spotted a recurring meeting that fell on a weekend and resolved it (something he had been putting off out of laziness), and it also detected a conflict between a DMV appointment and his travel plans. For the author, these are small but telling examples of what really interests him: that AI uses what it already knows to spot work it has not been explicitly asked to do.

Technically, a Dot runs on the GPT-6 Astra model, has its own computer in the cloud and can operate through connected apps. Its memory includes ChatGPT's memories, recent conversation context and information from connected apps, but it does not automatically inherit every previous chat: the user decides what to connect and which responsibilities to delegate to it. Nate explicitly values that limit.

The author compares this with Meta's Muse, which he notes also has its own computer, memory and a free entry route alongside paid options. He therefore concludes that asking 'does it have a computer and memory?' is a poor purchasing criterion; what matters is what specific responsibility the agent can take on, what it needs to know to do so and how much work still falls to the person to get the result.

Nate says he is especially interested in responsibilities that depend on substantial context accumulated with AI: projects already explained, assumptions already corrected, client needs already worked through. The long-term bet, he says, is that this understanding remains useful across conversations, avoiding having to rebuild the brief every time something changes. He cites the launch example OpenAI uses: when a product's scope changes, the Dot prepares revisions of the launch materials; in practice, he notes, a delayed feature can also affect the announcement, a sales document and promises already made to a client, and that is where the useful work begins: knowing which promises depended on that feature. The result, according to the author, is that the team arrives at its next conversation with the client with a precise proposal and a clear account of what changed, even though the final decision on what to promise is still made by a person.

On Space, Nate explains that it fits the same announcement because it lets colleagues and AI work on shared documents (Pages): a salesperson can flag a commitment made to a client while the product lead explains what has changed, and the next revision can draw on both contributions. The author observes that much of the current use of AI depends on a single person doing all the explaining (who knows the history, which answer is wrong, which exception invalidates the obvious solution), and that if the work passes to a colleague, much of that explanation has to be repeated. For him, this connects to the broader question about the AI-native workplace: how one person's reasoning helps the next person doing the same job. Shared pages give that reasoning somewhere to live, though the value depends on people making the relevant knowledge visible, including the why behind a decision.

As for availability, Space and Pages are available on the Pro, Business and Enterprise plans; collaborative spreadsheets and presentations are announced as coming soon. Business and Enterprise also add team tasks for recurring work and ChatGPT's participation in Slack and Teams. Nate clarifies a privacy nuance: sharing a page does not automatically share private chats or personal memory, but any content written on the page — including data pulled from memory — is visible to collaborators.

The author places this move on territory already contested by Microsoft, with Copilot and Autopilot, and Anthropic, with Claude Tag. As he reads it, Microsoft starts from employers that already buy its work environment, Claude Tag works with teams that are already on Slack, and OpenAI, by contrast, gives those who already work heavily inside ChatGPT a reason to bring their colleagues into that same app. His practical recommendation is to start from where the team's decisions, documents and people already are: moving the work has to pay off, and if ChatGPT already concentrates substantial project knowledge, Space becomes a concrete reason to try it; if the necessary information and permissions live elsewhere, the agent needs a reasonable route to reach them.

On GPT-6.1 Sol, Nate explains that as soon as recurring work is delegated to AI, cost stops being ignorable: an occasionally brilliant answer demands less of the plan than a task you want to repeat every day. In his experience, Sol has been one of the most useful pieces of this launch: he has used it for documents, code and research, and describes it as close to Astra in quality while consuming very little of his plan's quota. For heavy users of paid plans, he notes, that is usually the real cost question: how much work can be done before hitting the limit.

Beyond the editorial content, the email includes two benefits for paying subscribers: access to 'Nate's Library MCP', a read-only MCP server that connects the full archive of Nate's published articles directly to the reader's usual AI conversations, and access to his Slack community, where the author shares content between articles and helps in real time with projects ('builds').