GPT-5-Codex

What GPT-5-Codex is

GPT-5-Codex is a version of OpenAI's GPT-5 that the company further optimised for agentic coding. "Agentic" here means the model works through a task over many steps — reading files, editing code, running commands, checking the result — instead of producing a single answer in one pass. OpenAI described it as equally suited to quick interactive sessions and to long, complex tasks.^[5] It was released on 15 September 2025 and became the default model in Codex, the company's coding product.^[1]

Where it ran

Inside Codex, GPT-5-Codex became the default model for cloud tasks and for code review. Developers who wanted to work on their own machines could still select it for local work in the Codex CLI and the IDE extension.^[5] Codex was included with the ChatGPT Plus, Pro, Business, Edu and Enterprise plans, and OpenAI said its code review could catch critical bugs before release.^[5] On 23 September 2025 OpenAI made GPT-5-Codex available to developers using Codex with an API key, at GPT-5 prices.^[5]

API specifications

The model reached the API in September 2025, and only through the Responses API.^[2][4] Its context window is 400,000 tokens, of which 272,000 can be input, with up to 128,000 output tokens; it accepts text and image input, and its knowledge cutoff is 30 September 2024.^[2] Pricing is $1.25 per million input tokens and $10 per million output tokens, the same as GPT-5.^[2] It supported streaming, structured outputs, function calling, web search and prompt caching.^[2]

Why it mattered

Coding is one of the clearest cases where a model has to act rather than simply answer. A program that spans many files forces the model to hold earlier decisions in view, keep track of what it has already changed, and recover from its own mistakes — which is why a large context window and a long output allowance are useful for this kind of work rather than incidental.^[2] GPT-5-Codex was presented as exactly that sort of working model: the same GPT-5 base, tuned for the agentic setting, and priced identically to the base model.^[2][5] Charging the same rate made the choice between the general and the coding variant a question of fit rather than of budget.^[2]

The release also shows how OpenAI arranged the Codex product around a single default: one model for cloud tasks and code review, the same model available to pick locally, and the same model reachable through the API with a key.^[5] Rather than offering many specialised endpoints, the line-up centred on one working model that users could reach from several surfaces.

Lineage and retirement

GPT-5-Codex was built on GPT-5, and in November 2025 it was succeeded by GPT-5.1-Codex and GPT-5.1-Codex-Max.^[1] OpenAI deprecated gpt-5-codex on 22 April 2026 and shut it down on 23 July 2026.^[3] The short interval between launch and shutdown is a reminder that a named model version is a fixed snapshot: once a successor ships, the older identifier stays available only for a while before requests to it stop working, so anything built on a specific version has to plan for its replacement.