OpenAI releases GPT-5-Codex for its Codex agent

Background

Agentic software engineering describes a way of using a model that goes beyond answering questions in a chat window: the model works inside a code repository, reads and edits files, runs commands and tests, and keeps going over many steps until a task is finished. That demands something different from ordinary code completion — the model has to stay coherent across a long session and to call tools reliably again and again.

Codex is OpenAI's product for this kind of work, and two weeks before this release it had been unified into a single product experience, so that work could move between the local environment and the cloud without losing context.^[4] That unification is the setting in which the new model arrived.

What was released

On 15 September 2025 OpenAI released GPT-5-Codex, a version of GPT-5 tuned for agentic software engineering.^[1] "Tuned" here means a member of the same GPT-5 family whose behaviour has been adjusted toward a particular kind of work, rather than a general-purpose conversational model.

It became the default model for the Codex agent.^[1] A default matters because it is the version most users of the agent actually get, without having to choose a model name or setting; separate API access is what lets other software use the same model inside its own tools.

The model was not shipped on its own. It came as part of a wider update to Codex that also covered the command-line interface (CLI), an extension for integrated development environments (IDE), cloud tasks and code review.^[4] Those are the places where coding agents are actually used: the CLI for working in a terminal, the IDE extension for working inside an editor, cloud tasks for running work on remote machines rather than the developer's own computer, and code review for commenting on changes before they are merged.

API access and pricing

Developers who wanted to call the model directly had to wait a little longer. API access followed later in September 2025, at GPT-5 prices of $1.25 per million input tokens and $10 per million output tokens.^[2] Tokens are the units models read and write — roughly word fragments — so the two numbers set the cost of sending work in and getting code back out, with input charged at one eighth of the output price.

On 23 September 2025, GPT-5-Codex became available to developers using Codex via an API key, and only in the Responses API.^[4] The restriction matters for anyone integrating the model: the Responses API is one of the interfaces OpenAI offers for calling its models, and a model offered only there is not reached in the same way through other interfaces.

Reaction

The release landed in a contest between two coding agents. WIRED reported that Codex usage grew sharply in the months that followed, relative to its main rival, Anthropic's Claude Code.^[3]

Read together, the details describe a model released less as a standalone artifact than as a component of a product: shipped first as the default inside Codex, extended across the CLI, an editor extension, cloud tasks and code review, and only then opened to developers through the API, at the same token prices as GPT-5.