OpenAI launches Codex: an AI agent that writes code in place of engineers

OpenAI has released a cinematic video, "It's time to fly," a signal that Codex, its flagship AI agent for software development, has moved from beta tool to full-fledged product.

What Codex can do in 2026

Codex is built on GPT-5.5, the first fully retrained base model from OpenAI since the release of GPT-4.5. The model was trained with agentic scenarios in mind: it checks its own code before handing over the result, and it can chain hundreds of sequential tool calls without human intervention.

The key new feature is parallel work in several sandbox environments at once:

• runs a test suite in one environment
• generates documentation from a diff in parallel
• proposes a refactoring in a third sandbox
• reads the screen and controls GUI applications to automate QA
• integrates into the CLI, IDE, ChatGPT and GitHub under a single account

While the engineer reviews the results, Codex is already working on the next task.

The numbers and the scale

The product's launch did not go unnoticed. By the time the promo came out, Codex had more than 4 million active developers per week. NVIDIA gave 10,000 of its employees access to the agent, and not only engineers but also non-technical teams.

Harness, a DevOps company, published a case study on the OpenAI website about how Codex is moving from assistant to full participant in the engineering process: an "agent-first world" where the agent closes the task rather than suggesting how to close it.

What this means

The slogan "It's time to fly" describes the shift accurately: Codex is ceasing to be a smart autocomplete and becoming an autonomous executor. For development teams, this means routine engineering tasks (tests, documentation, small refactorings) are increasingly delegated to the agent, while people switch to architectural decisions and review.