OpenAI Launches Cloud-Based Codex and Computer-Use Agents API at DevDay 2026

OpenAI has granted software developers the immediate ability to bind artificial intelligence models to live desktop environments and cloud infrastructure. Through the new Computer-Use Agents API and Cloud-Based Codex released at DevDay 2026, engineering teams can now authorize autonomous systems to move cursors, click interfaces, and execute code on their behalf.[1][4]

The rollout fundamentally changes what developers are building with the company's models this quarter. Instead of prompting a chat interface for code snippets, enterprise teams can now deploy agents that navigate legacy software and write directly to cloud environments.[3]

"We are moving from models that talk to models that do," OpenAI stated during the DevDay 2026 keynote presentation.[4]

The company confirmed that the Computer-Use Agents API is currently available in a restricted beta for developers on the new $500 monthly enterprise tier, which includes 50,000 monthly agent actions. Broader access is scheduled for late November 2026, pending the resolution of early security audits.[3]

The Computer-Use Agents API

The most significant capability shipped this week is the Computer-Use Agents API, which bridges the gap between text generation and graphical user interface control. The system translates natural language commands into specific screen coordinates and keystrokes.[2][6]

According to documentation published by LMS Pedia, the API allows developers to build applications where an AI agent can open a web browser, navigate to a specific URL, and extract data without human intervention. The system relies on a continuous loop of screen capture and action prediction.[6]

"The agent essentially looks at a screenshot, decides where the mouse needs to go, and issues the command," notes independent developer Simon Willison in his live blog of the event.[5]

Willison pointed out that while the demonstration looked seamless, the actual latency of sending screen states to a cloud model and waiting for a coordinate response remains a practical hurdle for real-time applications.[5]

To mitigate the risks of runaway agents, OpenAI has implemented strict rate limits of 120 actions per minute and a mandatory human-in-the-loop confirmation step for destructive actions. Developers must explicitly whitelist which applications the agent is permitted to interact with.[1][2]

Cloud-Based Codex and Environments

Alongside the desktop control features, OpenAI overhauled its code-generation offering with the launch of Cloud-Based Codex. The platform now provides isolated, ephemeral execution environments directly hosted on OpenAI's infrastructure, allocating up to 16 gigabytes of RAM and 4 virtual CPU cores per sandbox.[7]

Developer Szymon Paluch reports that the new Codex CLI allows engineers to spin up a secure sandbox, prompt the model to write a Python script, and execute it immediately to test for errors. This eliminates the need to copy and paste code between a browser window and a local terminal.[7]

"The integration of automated code review directly into the cloud environment is the actual game-changer here," Paluch wrote following the DevDay announcements.[7]

The system includes automated security scans that evaluate generated code against a database of over 15,000 known vulnerabilities before execution. The Decoder noted that this feature, combined with the new Decisions API, allows agents to evaluate multiple execution paths and select the most secure option.[2]

However, the marketing language surrounding these releases often outpaces the shipped reality. While OpenAI promotes "Ultrafast" execution speeds, early testers report that complex multi-agent workflows still require 15 to 30 seconds of processing time and often fail on edge cases.[2][5]

Enterprise Pricing and Dots

To monetize these intensive workloads, OpenAI introduced a new $500 monthly subscription plan aimed at mid-market engineering teams. This tier provides higher rate limits and priority access to the new agentic endpoints.[3]

The company also unveiled "Dots," a new abstraction layer designed to help developers manage state and memory across long-running agent sessions. Reworked highlighted Dots as a critical infrastructure piece for enterprise customers trying to build reliable multi-step workflows.[3]

"Dots solves the amnesia problem that has plagued agent development for the last two years," the Reworked analysis concluded.[3]

By maintaining context over days rather than minutes, Dots allows a Computer-Use Agent to pause a task, wait up to 72 hours for human approval, and resume exactly where it left off without requiring a full prompt injection.[3][6]

Security and Ecosystem Impact

The shift toward autonomous execution introduces massive security implications, as a single agent can execute up to 1,000 interface interactions per hour. Granting an API endpoint the ability to click through a desktop environment bypasses traditional application programming interfaces and their built-in access controls.[1][2]

InfoQ's recap of the developer conference emphasized that security teams are already scrambling to establish frameworks for auditing agent behavior. The ability of an AI to interact with legacy software that lacks modern API security is both the primary selling point and the largest risk.[1]

As developers begin integrating the Computer-Use Agents API and Cloud-Based Codex into their production pipelines this month, the industry will test whether these tools can reliably automate rote engineering tasks.[4]

The true test will not be whether the agents can write code or click buttons in a controlled demonstration, but whether they can recover gracefully when a user interface changes or a cloud environment throws an unexpected error.[5][7]