OpenAI arrived at DevDay 2026 with more than two dozen announcements, but behind the sheer volume of updates, the direction is quite clear: the company wants artificial intelligence to stop being merely a tool we turn to for answers and increasingly become a technology capable of maintaining context, collaborating with teams, and carrying out work persistently.
The announcement that best represents that shift is Dots, a new generation of AI agents that can keep working toward goals and use different applications to move them forward. They are joined by ChatGPT Space, designed for collaborative work between people and agents; GPT-6.1 Sol, a new addition to its model family; the Decisions API; and new capabilities around Codex.
Rather than a collection of standalone products, DevDay reveals the different building blocks OpenAI wants to use to create a new way of working with artificial intelligence.
Dots was one of the standout stars of DevDay. The concept begins with an important difference from the traditional use of a chatbot: these agents are designed to remain active, retain the context of a project, and continue working toward an objective, instead of waiting for the user to initiate each new interaction.
Dots runs on GPT-6 Astra and can draw on other OpenAI tools, such as Codex and ChatGPT Work, depending on the task they need to solve. They can also interact with different applications and communicate with users through workplace tools such as Slack or Microsoft Teams.
During DevDay demos, OpenAI showed how one of these agents could review information coming from email and Slack, detect pending tasks, and later handle an API migration process. At first glance, that may seem like a small thing, but it changes the interaction model. Until now, much of generative AI usage has started with a person saying what they need. With a persistent agent, the goal is for AI to be able to follow a project’s development, recognize changes, and keep working without every step necessarily having to begin with a new prompt.
The fact that an agent can keep working inevitably raises another question: what can it do on its own, and what requires authorization? In an effort to define those boundaries, OpenAI added controls that make it possible to determine which actions a Dot can perform and which require approval, including especially sensitive operations such as changing passwords or permanently deleting information.
This becomes especially relevant after OpenAI decided to pause the release of a version of GPT-6.1 Astra after detecting, during internal evaluations, issues related to instruction-following, authorization boundaries, and the way the model reported certain actions.
In other words, the race to build more autonomous agents is also forcing AI companies to answer a less flashy but fundamental question: how to delegate work without losing control over important decisions.
If Dots suggests a new relationship between a person and their agent, ChatGPT Space extends that logic to teamwork, as the tool makes it possible to create shared spaces where several people can work alongside AI agents on the same project, while preserving information and context within the environment.
The proposal addresses one of the challenges emerging as AI becomes part of business operations: today, it is not enough for each employee to be able to use a chatbot to write, research, or analyze information. We also need to figure out how to integrate these systems into processes involving different people, tools, and sources of information.
In that scenario, an agent stops functioning solely as a personal assistant and begins to become another participant in the workflow. And that is precisely the layer where OpenAI appears to want to compete: not only by having the model people ask questions to, but by becoming the environment from which an ever-growing share of digital work is carried out.
OpenAI also introduced GPT-6.1 Sol, a new addition to the GPT-6 family aimed at offering a more cost-effective alternative for certain types of tasks; an announcement that comes just days after the introduction of GPT-6 Sol and GPT-6 Luna, launched on September 22.
Beyond the succession of names, the strategy behind this expansion is significant: OpenAI is building an offering in which companies can choose among different combinations of capability, speed, and cost, depending on what they want to solve. That becomes especially important as models stop being used only for one-off queries and begin operating within agents and processes that can execute hundreds or thousands of tasks.
Among the updates aimed at companies and developers is also the Decisions API, a tool designed to automate repetitive decisions within business processes. Perhaps one of the least flashy announcements at DevDay, it is also one of the clearest indicators of where enterprise AI usage is heading.
In marketing and eCommerce, there are countless processes in which an organization must continuously evaluate signals to decide what to do next: classify information, prioritize tasks, select an action, identify exceptions, or determine which cases need human intervention. But the possibility of embedding models within that layer opens a very different field from content generation.
The question would no longer be only “what can AI create?” but also “which repetitive decisions can we delegate to it, and under what rules?”
DevDay also brought new capabilities related to Codex, OpenAI’s programming-focused agent, as part of a strategy aimed at increasingly turning it into a collaborator capable of taking on full development processes, rather than simply generating snippets of code.
This evolution connects directly with Dots, because if an agent encounters a task during a project that requires developing or modifying software, it can rely on Codex to carry it out.
And that is where the architecture OpenAI is building begins to come into clearer focus: models with different levels of capability, persistent agents that maintain goals, specialized tools that carry out specific jobs, and spaces where humans and agents can collaborate.
Not every DevDay announcement will have the same impact or make its way immediately into everyday use, but taken together they help explain why OpenAI is placing so much emphasis on agents.
Over the past few years, our relationship with generative artificial intelligence has been shaped by a very specific dynamic: we write an instruction, receive a response, and decide what to do next.
DevDay’s announcements suggest a different kind of relationship, one in which AI can preserve the context of a project, use different tools, collaborate with other people, execute actions, and keep working toward an objective.
For companies, that also changes the conversation around AI adoption, because it is no longer enough to ask which model to use or which tasks can be accelerated with a chatbot. New questions are beginning to emerge about which processes can be delegated, what information an agent needs in order to operate, what permissions it should have, and at which moments a person must necessarily step in.
And perhaps that is the most interesting signal from DevDay 2026: the next stage of AI does not seem to be only about getting it to respond better, but about deciding how much work we are willing to place in its hands.
Your email address will not be published. Required fields are marked *
Δ