OpenAI DevDay Makes Agents Its New Platform Bet

OpenAI used its September 29 DevDay in San Francisco to make a case for a different kind of AI product: one that does not wait for a prompt, but can take on recurring work, use connected tools and continue across sessions. The company said it unveiled more than 20 announcements across ChatGPT, Codex, its models and developer tools. The clearest theme was not one new model; it was an attempt to turn ChatGPT into a platform for agents, software and shared work.

The headline consumer feature, Dots, is an always-available assistant that can carry a project across ChatGPT, Slack and Microsoft Teams, check connected information and perform agreed tasks. OpenAI says users can set boundaries, review progress and decide which apps a Dot may access. Its background “proactive research” is restricted to read-only connected tools; actions that could alter accounts or share information are subject to review and approval rules. These controls matter because a persistent agent is useful only if people can see what it is doing and stop it.

Agents move from demos into products

Dots are rolling out to Pro and Business Premium subscribers in eligible markets. Enterprise, Education and Healthcare customers can try a beta when an administrator enables it. OpenAI says a first Dot is included with eligible plans, but deeper work has plan limits, and access to apps or actions depends on permissions. The launch therefore describes a product direction, not a universal digital worker available to every ChatGPT user.

For developers, the Agents API adds computer use: an application can direct an agent through a hosted browser to interact with websites and software. OpenAI says the API also brings in tool search, tool calling, context compaction and multi-agent capabilities. In practice, this expands what a model can do beyond returning text or invoking a neatly defined API. It also introduces new failure modes: an agent can misread a screen, encounter hostile instructions or take the wrong action. OpenAI’s documentation calls for website-access approvals and for applications to handle sign-in; consequential results still need checking.

The separate Decisions API is narrower by design. It accepts text or images and returns typed outputs such as a choice from a set of options, a probability or a score. OpenAI positions it for classification, routing and other bounded decisions, and says it can respond up to ten times faster than the Responses API. As of October 12, it is in public beta and supports GPT-6 Luna. That makes it a specialized routing tool, not a general-purpose replacement for a human decision-maker.

A cheaper model and a cloud-based coding agent

GPT-6.1 Sol was another significant release. OpenAI describes it as nearing GPT-6 Astra performance on agentic coding, computer use and professional tasks, at one-fifth of Astra’s standard input and output token prices. The published API rates are $2 per million input tokens and $10 per million output tokens, compared with $10 and $50 for Astra; cached input is $0.10 for Sol and $1 for Astra. Those are meaningful differences for developers running repeated or long workflows, but the performance comparisons are OpenAI’s own evaluations, not an independent finding that the models are interchangeable.

DevDay also pushed Codex beyond a single laptop. Developers can run tasks in the cloud, connect from another device and reuse shared development environments. The refreshed command-line interface adds voice task steering and an agents view for tracking delegated work. Codex code review and Codex Security Cloud are aimed at reviewing code changes and scanning repositories. OpenAI’s October 8 API changelog later listed GPT-6.1 Sol Ultrafast as available to API users subject to rate limits, extending the event’s emphasis on speed and agentic work.

ChatGPT as a software destination

OpenAI is also opening ChatGPT to more third-party software. Plugin extensions can create interactive panels and file viewers inside the ChatGPT interface, while proposed Model Context Protocol events can trigger automations when something changes in a connected app. ChatGPT Space is a shared workspace for people, ChatGPT and Dots; Pages and collaborative slides extend the same idea to documents and presentations.

TechCrunch described the package as a challenge to the conventional app-store model: if people can discover and use services inside ChatGPT, the interface becomes a new distribution channel. OpenAI says ChatGPT has 1.2 billion weekly users, a company-reported figure that indicates why developers may pay attention. But the approach also puts OpenAI between users and the services they rely on. The practical questions include who controls app discovery, what data an extension can access, how permissions are explained and whether a business can move its workflows elsewhere.

What the announcements do—and do not—show

The event’s significance lies in the way these releases fit together: a lower-cost model, persistent assistants, remote coding, computer use and a route for third-party apps to live inside ChatGPT. OpenAI is competing not only on model quality but on the environment in which work gets assigned and completed. Axios and InfoQ likewise highlighted agents, Codex, the API and shared workspaces as the central direction of the event.

Availability remains uneven. Some features are in beta, limited to certain subscriptions or markets, or described as coming soon. OpenAI’s demonstrations and benchmark claims show its product ambitions, not proof that autonomous agents can reliably manage complex work without supervision. Businesses adopting them will need to test the systems against their own tasks, set narrow permissions, keep an audit trail and require human review for high-impact actions. DevDay made the agent-first strategy unusually explicit; its success will depend on whether the tools prove dependable, secure and worth the cost outside a keynote.