Evidence at a glance
The Product Is More Than a Conversational Assistant
At DevDay 2026, OpenAI introduced Dots, personal agents that can have names and independent identities. Dots can connect to services such as ChatGPT, Slack, and Teams, then use models, computers, and code execution environments under user authorization. In the Dottie demonstration, the agent did more than answer questions. It worked inside a shared collaboration space and asked Codex to build an application on a laptop and launch it in the iPhone Simulator. Access was opened that day to ChatGPT Pro and Enterprise customers, while OpenAI said it was also developing specialist Dots for areas such as legal and finance.
The important shift is that the product boundary has moved from asking for an answer in a chat window to delegating a piece of work to an actor with an identity. OpenAI has already placed ChatGPT, Codex, and development tools in different parts of the workflow. Dots attempts to pull them together. The competitive unit is therefore no longer just model output. It is the entire chain of identity, permissions, tool use, and task completion.
The Core Mechanism Is Identity Plus Execution
The design emphasis of Dots is to make an agent a recognizable participant in a team system. OpenAI employees have reportedly delegated tasks to agents directly in Slack, giving those agents their own identities. The agent is no longer merely a hidden feature behind an employee account. It can appear in message flows, shared workspaces, and delivered results. ChatGPT Space provides a shared container that resembles a document workspace with composable elements for organizing tasks, information, and agent actions.
The architecture links three kinds of capability. The first is model capability: GPT-6.1 Sol is positioned as a lower-cost option for routine reasoning, while Astra handles more demanding work. The second is tool capability, including OpenAI’s internal tools, Codex, and computer-use environments. The third is distribution, allowing results to reach users through ChatGPT, Slack, Teams, or enterprise services. For an enterprise, the central questions are no longer simply whether the model is smart. They are who each agent represents, what it can read, what it can change, and who takes over when it fails.
Cost and Latency Are Becoming Agent Architecture
The model announcements explain why this product can exist economically. OpenAI described GPT-6.1 Sol as close to Astra in intelligence at roughly one-fifth of the price. At the same time, Ultrafast can reach up to 300 tokens per second, about eight times the speed of the standard version, but costs six times as much. It is being made available across the API, ChatGPT, and Codex. The Pro 500 plan includes Ultrafast and offers 25 times the usage of Plus, while the $200-per-month plan is being reopened to new subscribers.
This is not simply a larger model lineup. It creates different economic tiers for agent work. Background tasks such as Slack coordination, document organization, and code maintenance can favor the cheaper Sol. Ultrafast may be justified when a user is waiting, when an agent must make a rapid decision, or when the task has clear commercial urgency. OpenAI also previewed a Decisions API that lets the Luna model choose from predefined options in a fraction of a second. That is another example of trading generality for speed and predictability.
From Dots to Sites, Work Starts Becoming a Product
Another thread at DevDay was the direct conversion of agent output into publishable interfaces. ChatGPT Sites began as an internal prototype by OpenAI employee Jon Abrams. It later launched and reached 8 million hosted sites, with OpenAI saying that 70 percent of its employees are creating their own Sites. In the live example, a release-information table was turned into a Launch Radar application, while Codex continued monitoring changes to the shared table and updating the site. Sites also added SQLite databases and scheduled tasks that day.
This moves the platform beyond having an agent perform an operation for an employee. It lets the agent turn the result of that operation into an accessible product. Internal teams may find it easier to publish small dashboards, directories, and operational tools, with agents handling data preparation, code changes, and maintenance. For developers, OpenAI is offering more than model access. It is combining models with plugins, login, user context, and distribution. The closer these systems get to real products, the less sufficient generation quality becomes as a metric. Teams must also ask whether updates are correct, whether tasks can run twice, and who has final publishing authority.
The Live Demo Exposed the Unfinished Responsibility Boundary
The live demonstration did not proceed perfectly. Dottie entered a “still checking” state, followed by an awkward silence, and the voice interaction showed signs of failure. The live notes also recorded that the author had intended to use Codex Cloud to build a photo system for the blog on a phone, but switched to Claude Code for web after running into problems. These details do not prove that any product is broadly unusable. They do show that the central bottleneck for long agent chains is still dependable completion, not merely starting a task.
A model described as highly aligned is not automatically an agent that can be trusted with broad responsibility. Once a Dot can access Slack, Teams, a laptop, and a code environment, permission isolation, action auditing, pause controls, and human takeover become part of the product itself. Technical leaders can delegate repetitive coordination, internal-tool maintenance, and low-risk code changes, but should first divide work into reversible steps, grant each connection the minimum necessary permissions, and define human gates for timeouts, duplicate execution, and incorrect publication. Ultrafast points to the same conclusion through pricing: speed matters when it reduces waiting or prevents loss. Otherwise, a cheaper model and a narrower decision path may be the better architecture.