DevDay showed OpenAI wants to own the agent’s entire workflow

OpenAI used DevDay 2026 to present more than 20 announcements across ChatGPT, Codex, models, APIs, security and builder tools. The individual items are at different stages of maturity, but together they describe an agent with its own environment, application connections and a route into enterprise customers.

More than 20 announcements assemble one agent platform

OpenAI’s primary page was blocked during verification, so this account relies cautiously on its public metadata and corroborating coverage rather than inaccessible detail. The confirmed scope includes Dots, GPT-6.1 Sol, ChatGPT, Codex, APIs, security and additional tools for builders.

Dots are intended to operate as persistent agents with a cloud computer and connected applications. OpenAI also showed remote Codex work, shared workspaces, agent interfaces and new distribution routes. DevDay therefore looked less like a single model release and more like a catalogue of the layers required for a model to accept work, execute it and return it for review.

Developers are buying an operating layer, not only tokens

For product and engineering teams, the build versus buy decision is changing. Managed environments, sessions, tool connections and task recovery can remove infrastructure work, but they also move more operational logic to OpenAI. Model pricing captures only part of that commitment.

The connection among Codex, agent APIs, identity and distribution is the more consequential story. If applications share an account, permissions and budget across ChatGPT and partner services, OpenAI controls both the user entry point and the route through which an agent reaches third party software. That position is stronger than simply selling inference.

A broad catalogue conceals the gap between product and promise

More than 20 announcements do not equal more than 20 finished products with identical availability. Public reports distinguish live features, gradual rollouts, betas, limited previews and capabilities promised for later. Dots also begin in selected paid plans and, according to launch reporting, were unavailable in Europe and the United Kingdom at release.

Autonomous work expands the room for permission errors, prompt injection and irreversible actions. Custom rules, approvals and automated review are useful safeguards, but their presence is not evidence of reliability inside an unfamiliar enterprise environment.

Auditing and exit options will decide the winner

Teams should watch whether they receive complete action logs, explicit permission boundaries, reproducible failures and a practical way to move workflows elsewhere. It will matter just as much how many announced layers graduate from preview to stable service with documented regional availability.

DevDay made the direction clear: OpenAI wants to serve the entire agent work cycle. Operations will reveal whether that becomes a dependable platform or an expensive corridor where one vendor holds every access card.

Lilith's verdict

OpenAI is building more than another model booth. It is drawing the whole floor plan and wants to issue every access card itself.