DevDay 2026 Was About Distribution, Not Benchmarks
On September 29, 2026 OpenAI used its developer conference to ship an autonomous agent called Dots and an enterprise storefront named OpenAI Marketplace, alongside rebuilt plugins, a shared login and a faster review pipeline. The through line was not another model release. It was a plan to convert a product with roughly 1.2 billion weekly active users from a chat box into the place where software is found, launched, run and paid for.
Dots: An Agent That Keeps Working After You Stop Typing
Dots runs on GPT-6 Astra and gets its own cloud computer and browser. It connects to applications the user has authorized and continues executing long after a conversation ends. Instead of re-explaining context on every prompt, it holds a standing objective, advances the project, tracks dependencies, watches for failures, and can perform read only background research while nobody is actively chatting with it.
The connection surface already spans more than 4,000 applications and is reachable through ChatGPT, Slack and Microsoft Teams. Autonomy stays bounded: users define the access scope, choose which actions require approval, and can forbid specific operations entirely. Access is rolling out gradually to qualifying ChatGPT Pro, Business Premium and Enterprise accounts.
Plugins Became Applications Inside the Chat Window
OpenAI also reworked what a plugin can be. Developers can now build sidebar entries, interactive panels, editors, dashboards and full workspaces that live inside ChatGPT, so a task no longer requires hopping to another tab. Design files from Figma and Adobe tools appeared embedded in the product during the keynote, letting users review, edit and resolve comments without leaving the conversation. ChatGPT also surfaces relevant apps based on the task the user is working on.
The other structural change is Sign in with ChatGPT. Users can authenticate into third party products with their ChatGPT account, and eligible Plus and Pro members can carry their existing ChatGPT usage allowance into partner applications. Around sixteen platforms came up first, including Devin, Notion, Vercel and OpenClaw. That arrangement cuts both ways: developers are no longer always forced to absorb model inference cost on a new user, and users can try a third party AI product without a fresh subscription decision.
Marketplace Redirects Enterprise Budget to Partners
OpenAI Marketplace opened with more than thirty partners, among them Salesforce, HubSpot, CrowdStrike, Figma, Adobe and ServiceNow. Its mechanic differs from a conventional app store. Enterprises can direct part of the spend already committed to OpenAI toward partner software. In effect the company is no longer only selling its own models and seats; it is routing customer budgets into an ecosystem.
For independent developers that is a new distribution channel with an enterprise buyer already in the room. For customers it can compress multi vendor procurement. What remains unspecified is the money: OpenAI has not published a revenue share or billing structure comparable to the systems Apple and Google operate.
Infrastructure Built for Agents
Underneath the storefront work, the developer stack moved in the same direction. Plugin review became observable, with progress status, issue submission, human appeal, and the ability to ship updates without restarting an approval queue from zero. The Agents API gained Computer Use, so builders can create agents that drive software interfaces directly for testing, automation flows and cross application tasks. Codex pushed further into the cloud, letting long jobs continue running when the local machine is shut down.
Four Layers, and the Question They Raise
Read together, the announcements form a stack rather than a feature set. A unified identity layer lets accounts and usage allowances travel across applications. A unified entry point makes ChatGPT the surface where software is discovered and invoked. A unified execution layer puts a cloud computer behind an agent that acts across apps. A unified commercial layer lets enterprise budgets flow to third parties.
That is not the same as displacing mobile app stores. Downloads, native binaries, OS permissions and corporate procurement remain complicated, and OpenAI has no published payout model yet. But the direction is explicit, and it points toward an interaction pattern where a person states an intent, an agent selects the tooling, and execution follows automatically. The open question is how much control an operator wants to hand to software that keeps running after the conversation closes.
