Dots, ChatGPT Space, Codex, and new models all point in the same direction: AI needs to move from answering questions to taking over tasks.

Dots: AI Starts Watching Projects in the Background

On September 29, OpenAI released more than 20 updates at DevDay. The list is long, but the direction is highly focused: turning AI from something that answers one question at a time into software that can keep pushing tasks forward in the background. The clearest example of this shift is Dots.

Dots has a cloud computer and can use connected apps to handle work. After a user gives it a goal and sets the scope within which it can act on its own, it can keep running. The scenarios OpenAI demonstrated include maintaining sales proposals, updating materials as customer needs change, and generating demo products for review. These demos show that it can complete the relevant workflows, but they cannot answer questions about stability or rework volume during long-running operation.

Currently, Dots is available to Pro and Business Premium users, and some enterprise, education, and healthcare workspaces can enable testing through administrators. Access is still limited, and a product demo cannot prove that it can already reliably take over a job. For now, the more accurate judgment is that OpenAI has begun offering persistent agents as a product form for users to try.

OpenAI's official demonstration of the Dots enterprise proposal workflow. The image comes from a product demo and cannot prove long-term operating performance.

ChatGPT Space: Team Work Gets a Shared Context

If Dots is responsible for continuous execution, ChatGPT Space is more like a shared space for team work. Members can collaborate with ChatGPT inside it, turn conversations into pages or presentations, and invite colleagues to add or revise them. OpenAI also launched team tasks, @ChatGPT in Slack and Microsoft Teams, and meeting notes, among other features.

This shifts the focus of competition in enterprise software beyond how well the model answers questions to whether project materials can be retained, whether members can continue collaborating, and whether tasks can be delivered continuously. Whether users come back often and whether teams are willing to pay will ultimately depend on whether the product can enter daily work workflows.

ChatGPT Space and Pages collaboration interface in OpenAI's official DevDay demo, September 29, 2026.

Codex and API: Handing Agent Execution Capabilities to Developers

Developers can run Codex locally, remotely, or in the cloud. Reusable development environments make tasks start faster and also make it easier for teams to share approved tools and permissions. The Agents API adds computer operation capabilities; the Decisions API limits the model to the questions and answers preset by the user, and can be used for classification, routing, or deciding an agent's next action. It is still in limited preview.

OpenAI's official Codex Cloud product interface, September 29, 2026.

Plugin extensions and MCP event triggers are also worth noting. Developers can put apps into the ChatGPT sidebar, and can also let external events start automated workflows. OpenAI says developers can distribute native experiences to ChatGPT users. If this mechanism is accepted by users, ChatGPT could become both a model entry point and a distribution channel for AI applications.

Model Price Cuts and a High-Speed Tier Correspond to Two Kinds of Bills

GPT-6.1 Sol is the focus of this model update. OpenAI says it approaches GPT-6 Astra in some coding, computer operation, and professional work evaluations, while the standard input and output token unit price is about one-fifth that of Astra. Sol has entered the API, Codex, and ChatGPT Work, but has not yet entered ordinary ChatGPT conversations. These evaluations and the judgment that it is "close to Astra" come from OpenAI and cannot be directly extended to all real tasks.

Ultrafast turns speed into a high-priced tier: Codex up to 300 tokens per second, and the API up to 6 times faster. The $500-per-month Pro 500 plan provides 25 times the usage of Plus. Product pricing thus splits into two ends: Sol lowers the cost of calling complex tasks, while Ultrafast and high-usage subscriptions target customers willing to pay for speed and volume. Public information is still insufficient to judge the gross margins and actual usage costs at either end.

After the Launch Event, Watch Whether Tasks Can Be Completed Reliably

OpenAI wants to turn ChatGPT into the entry point for work, a platform for agents to call tools, and a channel for developers to reach users. Reuters reported that OpenAI said Codex and ChatGPT Work have more than 35 million weekly active users, while ordinary ChatGPT has more than 1.2 billion weekly active users. These numbers show that it has a large distribution base, but user scale does not equal revenue, and API call volume does not equal profit.

Agents still have to pass the tests of reliability and permission management. Reuters recorded that the voice update demo at the DevDay site failed to return normally several times. OpenAI says users can set rules, and sensitive operations such as changing passwords or permanently deleting data require explicit consent. As agents can do more, completing tasks reliably, explaining the execution process, and stopping before overstepping authority will all affect whether users dare to hand work over to them.

What is worth tracking next is Dots' continued usage rate, how much manual rework is needed after tasks are completed, the model cost per job, and whether enterprise customers renew. These are closer to commercial answers than launch event demos. Before the data comes out, Dots represents a new product direction that is being tested, not a proven job replacement solution.

Sources: OpenAI DevDay 2026 official recap and product descriptions; Reuters, September 29, 2026 report. Performance, pricing, and active user data in the text are attributed according to the sources.