OpenAI today released a research preview of Codex, its most capable AI coding agent yet, integrated into ChatGPT. The agent runs in a sandboxed cloud environment and is powered by codex-1, a version of the o3 reasoning model optimized for software engineering. OpenAI says codex-1 produces cleaner code, adheres more precisely to instructions, and iteratively runs tests until passing results are achieved.

Codex can handle multiple software engineering tasks simultaneously, taking one to 30 minutes to write features, fix bugs, answer questions about a codebase, or run tests. It connects with GitHub to preload repositories. Starting today, Codex is available to ChatGPT Pro, Enterprise, and Team subscribers with generous initial access, though rate limits will be introduced in the coming weeks with an option to purchase additional credits.

The launch comes amid a surge in AI coding tools, often called vibe coding. Google and Microsoft CEOs claim roughly 30% of their companies’ code is now AI-generated. Anthropic released Claude Code in February, and Google updated Gemini Code Assist in April. Cursor, a popular AI coding tool, reached $300 million annualized revenue in April. OpenAI has reportedly closed a $3 billion deal to acquire Windsurf, another AI coding platform, signaling its aggressive push into the space.

OpenAI says Codex will reliably refuse requests to develop malicious software, operating in an air-gapped environment with no access to the broader internet or external APIs. However, the tool may still make mistakes: a recent Microsoft study found that leading AI coding models struggled to reliably debug software. Despite this, investor excitement in AI coding tools remains high, and OpenAI hopes Codex will entice more users to subscribe to ChatGPT and pay for additional usage.