How OpenAI uses Codex

Codex is used daily across numerous technical teams at OpenAI like Security, Product Engineering, Frontend, API, Infrastructure, and Performance Engineering. Teams are using it to accelerate a range of engineering tasks, from understanding complex systems and refactoring large codebases to shipping new features and resolving incidents under tight deadlines.

Drawing from interviews with OpenAI engineers and internal usage data, we’ve compiled use cases and best practices that highlight how Codex helps our teams move faster, improve work quality, and manage complexity at scale.

Codex helps our teams get up to speed quickly in unfamiliar parts of the codebase when onboarding, debugging, or investigating an incident.

They often use Codex to locate the core logic of a feature, map out relationships between services or modules, and trace data flow through a system. It also helps surface architecture patterns or missing pieces of documentation that would otherwise require significant manual effort to generate.

During incident response, Codex helps engineers ramp into new areas quickly by surfacing interactions between components or tracing how failure states propagate across systems.

Codex is commonly used to make changes that span multiple files or packages. For example, when engineers are updating an API, changing how a pattern is implemented, or migrating to a new dependency, Codex makes it easy to apply changes consistently.

It’s especially useful when the same update needs to be made across dozens of files, or when the update requires awareness of structure and dependencies that aren’t easily caught with a regex or find-and-replace.

They’re also using it for code cleanup by breaking up oversized modules, replacing old patterns with modern ones, or preparing code for better testability.

Codex is used to identify and address performance bottlenecks.

During tuning or reliability efforts, engineers prompt Codex to analyze slow or memory-intensive code paths, such as inefficient loops, redundant operations, or costly queries and suggest optimized alternatives, often resulting in meaningful gains in efficiency and reliability.

Codex is also used to support code health by identifying risky or deprecated patterns that are still in active use. Our teams lean on it to help reduce long-term tech debt and proactively prevent regressions.

Codex helps engineers write tests faster — especially in places where coverage is thin or completely missing.

When working on a bug fix or refactor, engineers often ask Codex to suggest tests that cover edge cases or likely failure paths. For new code, it can generate unit or integration tests based on the function signature and surrounding logic.

Codex is particularly helpful for identifying boundary conditions like empty inputs, max length, or unusual but valid states that are often missed in initial tests.

Codex helps teams move faster by accelerating both the start and end of the development cycle.

When kicking off a new feature, engineers use it to scaffold boilerplate — generating folders, modules, and API stubs to get runnable code up quickly without hand-wiring every piece.

As projects approach release, Codex helps meet tight deadlines by handling smaller but essential tasks like triaging bugs, filling in last-mile implementation gaps, and generating rollout scripts, telemetry hooks, or config files.

It’s also used to turn product feedback into starter code. Engineers often paste in a user request or spec and have Codex generate a rough draft they can return to and refine later.

Codex helps our engineers stay productive when their schedules are fragmented and filled with interruptions.
It’s used to capture unfinished work, turn notes into working prototypes, or spin off exploratory tasks that can be revisited later. This makes it easier to pause and resume work without losing context, especially when they’re on call or have a lot of meetings.

Codex is also useful for open-ended work like finding alternative solutions or validating design decisions. You can prompt for different ways of solving a problem, explore unfamiliar patterns, or pressure-test assumptions. This helps surface tradeoffs, expand design options, and sharpen implementation choices.

It’s also used to identify related bugs. Given a known issue or deprecated method, Codex can identify similar patterns elsewhere in the code, making it easier to catch regressions or finish cleanup work.

Codex works best when it’s given structure, context, and room to iterate. Here are some of the habits OpenAI teams are cultivating to get consistent value out of it in day-to-day work.

For large changes, start by prompting Codex for an implementation plan using Ask mode, which then becomes the input for follow-up prompts when you switch to Code Mode. This two-step flow keeps Codex grounded and helps avoid errors in its output. Codex works best with well-scoped tasks that would take you or a teammate about an hour to complete or a few hundred lines of code to implement. As models improve, expect the size of the tasks it can take on to increase.

Setting a startup script, environment variables, and internet access significantly reduces Codex’s error rate. As you run tasks, look for build errors that can be corrected in Codex’s environment configuration. This may take a few iterations, but gives significant efficiency gains in the long run.

Codex responds better when prompts mirror how you’d describe a change in a PR or issue. That means including file paths, component names, diffs, and doc snippets when relevant. Prompting with patterns like “Implement this the same way it’s done in [module X]” improves results.

Fire off tasks to capture tangential ideas, partial work, or incidental fixes. There’s no pressure to generate a full PR in one go. Codex works well as a staging area you can return to when you’re back in focus.

Maintain an AGENTS.md file to help Codex operate more effectively in your repo across prompts. These files typically include naming conventions, business logic, known quirks, or dependencies Codex can’t infer from the code alone. Learn more on structuring your AGENTS.md file in the docs.

The Best-of-N feature lets you simultaneously generate multiple responses for a single task to quickly explore multiple solutions and pick the best one. For more complicated tasks, you can review several iterations and combine parts of different responses to get a stronger result.

Codex is still in research preview, but it’s already making a real impact in how we build, helping us move faster, write better code, and take on work that would’ve otherwise never been prioritized.

We’re excited by the potential ahead — as our models get better and Codex becomes more deeply integrated into our workflows, we’re looking forward to unlocking even more powerful ways to develop software with it. We’ll continue to share what we learn along the way.