OpenAI Codex

OpenAI Codex is a software engineering agent developed by OpenAI that performs coding work on a user's behalf — writing features, fixing bugs, running tests and proposing pull requests — running either in cloud sandbox environments or locally through a command-line tool and IDE extension. The name was first used for a 2021 language model specialised in code; OpenAI revived it in May 2025 for an autonomous agent, and by 2026 Codex had become one of the company's flagship developer products and a centrepiece of the commercial market for AI coding agents.^[1][2][3]

History

The original Codex was a 2021 model derived from GPT-3 and fine-tuned on public code, best known for powering the initial version of GitHub Copilot; OpenAI retired the API model in 2023. The modern product launched on 16 May 2025 as a research preview: a cloud-based agent powered by codex-1, a version of the o3 reasoning model trained with reinforcement learning on real software tasks, available first to ChatGPT Pro, Business and Enterprise users and expanded to Plus subscribers in June 2025. Each task ran in its own isolated cloud sandbox preloaded with the user's repository, with completion typically taking one to thirty minutes and results delivered as reviewable commits and pull requests. An open-source command-line interface and an IDE extension followed, giving developers local ways to invoke the same agent. In September 2025 OpenAI introduced GPT-5-Codex, a model tuned for long-horizon agentic coding and code review, and through 2026 the product evolved into a multi-agent workspace: GPT-5.3-Codex arrived in February 2026, with OpenAI describing an agent able to do "nearly anything developers can do on a computer", and a Codex desktop application let multiple agents run in parallel on separate copies of a repository. By September 2026 Codex was presented as a command centre inside ChatGPT — with reusable Skills for team conventions and scheduled background work such as issue triage and CI/CD tasks — running on OpenAI's current frontier models.^{[1][2][3][4][5][6]}

Key Concepts and Technology

Codex is an agent in the sense that it executes multi-step work rather than autocompleting lines of code: it reads a repository, plans changes, edits files, executes tests and iterates until checks pass, then reports what it did with citations to terminal logs and test output so a human can verify the result. Configuration files known as AGENTS.md let teams describe project conventions — how to build, test and navigate a codebase — which the agent follows consistently. The cloud model isolates each task in a disposable environment, enabling parallel workstreams and protecting the user's machine; the local CLI and IDE modes suit interactive use where a developer watches and steers. OpenAI emphasises human review of all generated code before integration, reflecting the industry's broader safety posture for coding agents, and charges for usage through ChatGPT subscription tiers with rate limits that have been revised several times, most notably in April 2026.^[1][2][6]

Adoption and Impact

Codex competes with Anthropic's Claude Code, Cognition's Devin, Cursor and GitHub Copilot, and its launch helped normalise the idea that substantial software work can be delegated to agents under review. OpenAI says Codex is used by engineering teams at companies such as Ramp, Duolingo and Cisco Meraki for pull-request review, migrations and test generation, and the February 2026 release was accompanied by the company's own account that earlier Codex versions had been used internally to write deployment tooling and evaluation scripts while building the successor model — an example of the recursive use of AI in AI development. Analysts debate how far the category substitutes for engineering headcount, but adoption has been broad: coding agents were among the fastest-growing enterprise AI segments of 2025 and 2026.^[3][4][5]

>See Also

• ChatGPT
• GPT-5
• GitHub Copilot
• Claude Code
• Devin
• Vibe Coding