OpenAI has launched a new AI coding assistant, Codex, built to help developers navigate and automate common software engineering workflows. The tool is designed to generate code, fix issues, and write tests — all with minimal input from users. It operates on the codex-1 model, a variant of OpenAI’s o3 model, customized specifically for engineering productivity.
What the Codex Agent Does
Codex functions as a built-in companion within ChatGPT, accessible via a side panel. Developers assign tasks using simple prompts, choosing between code generation or general queries about their codebase. Each request is processed in an isolated, cloud-based environment that mirrors the user’s project files.
Task duration varies depending on scope, with some completing in under a minute and others taking up to 30 minutes. Once finished, Codex provides a full audit trail — capturing what actions it took, outputs from the terminal, and any test results — so developers can review exactly what occurred.
To guide the agent’s behavior, teams can optionally include an AGENTS.md file in their project directory. This functions as a blueprint for Codex, detailing conventions, testing routines, and structural information. Even in the absence of custom instructions, Codex remains effective, but human oversight remains a key safeguard before integrating its output.
OpenAI also emphasized that Codex has been trained to recognize and block misuse, particularly attempts to create malicious code. Accompanying its release, the company has updated its safety documentation and evaluation criteria.
Early Enterprise Adoption
Prior to public rollout, OpenAI granted access to a handful of companies for early testing. Cisco, for instance, has explored Codex’s potential to streamline software development across its teams.
Additional early adopters include:
- Temporal, which has been using the tool to create tests and locate bugs more efficiently.
- Superhuman, where product leads can now apply minor code tweaks without involving developers.
- Kodiak, which is evaluating Codex to accelerate improvements to its autonomous vehicle systems.
Who Can Use It, and What It Costs
As of now, Codex is available to users on ChatGPT Pro, Team, and Enterprise plans. OpenAI said access will soon expand to users on Plus and Edu tiers. While use is currently free, the company will implement rate limits and introduce usage-based pricing soon.
For API users working with the codex-mini-latest model, pricing is structured as follows:
- $1.50 per million input tokens
- $6.00 per million output tokens
- 75% discount on repeated prompts (via caching)
Codex remains in preview, and certain limitations apply. It doesn’t yet support image-based tasks, and cannot be interrupted mid-task. In some cases, assigning work to Codex may take longer than direct coding. OpenAI said upcoming versions will improve interactivity, add mid-task guidance, and integrate more deeply with developer tools such as CI systems and issue trackers.
“Eventually, Codex will feel less like a tool and more like a remote teammate you collaborate with asynchronously,” the company said.
This article was based on reporting by contributor Aminu Abdullahi.