OpenAI has upgraded and released Codex, a new AI agent designed to help developers write and manage code more efficiently.
Powered by codex-1, a version of OpenAI’s o3 model tailored for software engineering tasks, the agent can simultaneously handle multiple coding jobs, from writing new features to fixing bugs and submitting pull requests for review.
Codex was first released on May 16 for ChatGPT Pro, Team, and Enterprise users. On June 3, OpenAI made Codex available for ChatGPT Plus users, too.
How Codex works
Codex lives inside ChatGPT and can be accessed through a simple sidebar. Developers assign tasks using prompts and select either “Code” (to generate new code) or “Ask” (to get answers about their codebase). Each task runs in a secure cloud sandbox preloaded with the developer’s codebase.
Internet connection for Codex is opt-in and available only for Plus, Pro, and Team users. Once the user selects internet access and enables it for specific environments, Codex can “install dependencies, upgrade packages, run tests that need external resources, and more,” OpenAI detailed on June 3.
OpenAI states that task completion takes between one and 30 minutes, depending on the complexity. Once the task is done, Codex logs and cites every action it took, including terminal outputs and test results, so users can verify what happened.
Developers can guide Codex using a file called AGENTS.md, which serves as a README to help the agent understand the project structure, testing commands, and preferred practices. However, OpenAI says codex-1 still performs well even without custom instructions, though human review is essential before integrating AI-generated code.
“Codex was trained to identify and precisely refuse requests aimed at development of malicious software, while clearly distinguishing and supporting legitimate tasks,” the company said in a statement.
OpenAI also updated its system documentation to reflect new safety evaluations and guidelines. The safety evaluations were updated when internet access was added. It noted that exposure to the online world invites “prompt injection, exfiltration of code or secrets, inclusion of malware or vulnerabilities, or use of content with license restrictions.”
For added safety, Codex’s internet access can be restricted using a domain allowlist. Users can either set allowed domains themselves or use a preset list of common sites for working with dependencies, such as GitHub and Google.
Cisco among early Codex users
Before its public release, a select group of companies received early access to test Codex in real-world settings. For instance, Cisco has been exploring how Codex can speed up development across its product lines.
Other early testers include:
- Temporal, which uses Codex to more quickly write tests and debug.
- Superhuman, where product managers can make minor code edits without pulling in engineers.
- Kodiak integrates Codex to improve its autonomous driving software.
SEE: How to Use AI in Business from TechRepublic Premium
Who has access to Codex?
Currently, Codex is available for ChatGPT Plus, Pro, Team, and Enterprise users. Access for Edu users is “coming soon,” according to OpenAI. The company is initially offering free usage, with plans to introduce rate limits and pay-as-you-go pricing in the coming weeks.
Pricing for developers using the codex-mini-latest model via the API is as follows:
- $1.50 per 1M input tokens
- $6 per 1M output tokens
- 75% prompt caching discount
Still in preview, Codex doesn’t support image inputs for frontend tasks and lacks real-time task interruption. Also, delegating tasks to Codex can take longer than working interactively. The company says that future versions will include more interactive workflows, mid-task guidance, and deeper integrations with tools such as issue trackers and CI systems.
OpenAI noted: “Over time, interacting with Codex agents will increasingly resemble asynchronous collaboration with colleagues.”
TechnologyAdvice staff writer Megan Crouse contributed to this article.