Introducing Codex is OpenAI’s announcement of a research preview of OpenAI Codex, published on 16 May 2025 and rolled out first to ChatGPT Pro, Enterprise and Business users. It presents Codex as a cloud-based software engineering agent that can work on many tasks in parallel — writing features, answering questions about a codebase, fixing bugs, and proposing pull requests — with each task running in its own cloud sandbox preloaded with the user’s repository.
The post is explicit that this is an early, iterative release rather than a finished product, and the version of the page read here carries a later notice stating that the launch post is outdated and pointing readers to the current Codex product pages. A dated update of 3 June 2025 records that Codex became available to ChatGPT Plus users and that internet access during task execution was being enabled.
Key Points
- Codex is powered by codex-1, described as a version of OpenAI o3 optimized for software engineering and trained with reinforcement learning on real-world coding tasks so that it produces code close to human style and PR preferences, follows instructions precisely, and can iterate on tests until they pass.
- Each task is processed independently in an isolated environment where Codex can read and edit files and run commands including test harnesses, linters and type checkers, typically taking between 1 and 30 minutes.
- Verifiability is presented as a deliberate design choice: Codex cites terminal logs and test outputs so each step can be traced, and explicitly communicates uncertainty and test failures. OpenAI states it remains essential for users to manually review and validate all agent-generated code before integration and execution.
- Codex can be guided by AGENTS.md files in the repository, though OpenAI reports codex-1 performing strongly on coding evaluations and internal benchmarks even without them or any custom scaffolding.
- At launch the agent ran with internet access disabled, limiting it to the code in the connected GitHub repositories and to dependencies pre-installed by a user-supplied setup script. OpenAI notes separately that this describes the launch configuration only.
- On abuse, OpenAI states Codex was trained to identify and precisely refuse requests aimed at developing malicious software while still supporting legitimate work that uses similar techniques, such as low-level kernel engineering, and that it strengthened its policy frameworks and safety evaluations to reinforce those boundaries.
- OpenAI reports its own engineers using Codex mainly to offload repetitive, well-scoped tasks — refactoring, renaming, writing tests — and names four companies from a small group of external testers it worked with before release; those accounts are relayed by OpenAI rather than independently reported.
- Based on what it learned from those early testers, OpenAI recommends assigning well-scoped tasks to multiple agents simultaneously and experimenting with different task types and prompts.
- Alongside the cloud product it announces a smaller version of codex-1, derived from o4-mini,
as the new default model in Codex CLI and as
codex-mini-latestin the API, optimized for low-latency code Q&A and editing.
Context
The closing argument is about where the industry is heading rather than about the product: OpenAI writes that pairing with AI tools has quickly become an industry norm, but predicts the asynchronous, multi-agent workflow Codex introduces in ChatGPT will become the de facto way engineers produce high-quality code, with real-time pairing and task delegation eventually converging into one workflow. It names the limitations it accepts for now — no image inputs for frontend work, no way to course-correct the agent mid-task, and the latency of delegating to a remote agent — and says guidance mid-task, collaboration on implementation strategies and proactive progress updates are planned. That is a vendor’s forecast about its own product category, not an independent assessment.
An appendix publishes the codex-1 system message so that developers can understand the model’s default behaviour, and much of it is the AGENTS.md contract stated as rules the agent follows: that such a file’s scope is the directory tree rooted where it sits, that more deeply nested files take precedence on conflict, that direct system, developer and user instructions outrank them, and that any programmatic checks the file specifies must be run after all code changes. The rest covers git conventions and a citation format for referencing file paths and terminal output in the final response.