OpenAI drops Codex, an autonomous agent that codes in the cloud

OpenAI just launched Codex in research preview, turning ChatGPT from a conversational assistant into an autonomous software engineering agent that works in the background.

Powered by codex-1—a specialized variant of OpenAI o3 tuned with reinforcement learning on real-world engineering tasks—Codex takes assigned work, executes commands inside a private cloud environment, and produces verifiable git commits.

A sandboxed developer on your repo

Instead of watching tokens stream character by character, you assign Codex a task from the ChatGPT sidebar and walk away. Tasks take anywhere from 1 to 30 minutes, running in a clean cloud container loaded with your codebase.

Inside this sandbox, the agent can read and modify files, execute linters, and trigger test suites. In its default launch setup, internet access is strictly shut off to prevent unintended external network calls. If a test fails, codex-1 iteratively tweaks the code and reruns the suite until it passes.

Every action comes with receipts. When Codex finishes, it provides citations linking directly to terminal logs and test outputs so human developers can trace its trail. Teams can also guide its behavior by committing an AGENTS.md file to their repo, establishing rules much like a README for bots.

CLI upgrades and lightweight models

OpenAI also rolled out codex-mini-latest for its open-source terminal companion, Codex CLI. This model is a specialized version of o4-mini tuned for low-latency coding tasks and terminal edits.

Developers can access codex-mini-latest via the Responses API at $1.50 per million input tokens and $6 per million output tokens, along with a 75 percent discount on cached prompts. Connecting the CLI no longer requires handling raw API keys, as users can log in directly with their ChatGPT credentials, unlocking $5 in credits for Plus members and $50 for Pro subscribers.

Codex is rolling out immediately to Pro, Business, and Enterprise users, with Plus and Edu access following shortly. Early pilot partners like Cisco, Temporal, Kodiak, and Superhuman are already using it to handle background bug fixing, test writing, and triage.

Why it matters

We think Codex signals the industry shift from autocomplete copilots toward asynchronous junior colleagues. For developers, the immediate value is not having an AI invent new architectures, but offloading low-context chores like drafting tests, updating dependencies, and handling refactors without getting pulled out of deep work.

OpenAI's smartest move here is leaning on verifiable terminal output. An agent that simply claims it fixed a bug is useless, but an agent that runs your own test suite inside a container and proves it passed changes the trust equation completely.

However, working asynchronously comes with friction. The lack of mid-task steering means a slight prompt misunderstanding could waste up to 30 minutes of container time. As coding agents evolve, competitors will now be forced to move beyond IDE text suggestions and offer robust, isolated execution sandboxes.