GPT-5-Codex can grind for 7+ hours on a single bug fix

OpenAI just dropped GPT-5-Codex, a specialized branch of its latest model that is strictly tuned for agentic software engineering. This isn’t a general-purpose chatbot shoehorned into a coding window. The company trained it specifically on the messy, real-world stuff: building projects from zero, adding tests, performing massive refactors, and conducting deep code reviews. It is now the default brain inside Codex for cloud tasks and code review, and it’s available via API for developers who prefer to integrate it directly into their own pipelines, priced identically to standard GPT-5.

What separates this from a typical code-gen model is the runtime. OpenAI observed the model working independently for spans exceeding seven hours, iterating on implementations, fixing test failures, and only stopping when the job was done. The team baked in dynamic compute scaling to make this viable. The model now adjusts how hard it thinks based on the complexity of the prompt. For quick, interactive turns, it’s actually cheaper—on the bottom 10% of simple requests, GPT-5-Codex uses 93.7% fewer tokens than its general-purpose sibling. But when you throw a massive refactor at it, the meter runs harder, with the model spending twice as long reasoning, editing, and testing compared to standard GPT-5.

Code review is a standout feature. The model navigates the full codebase, reasons through dependencies, and even executes code to validate correctness before flagging an issue. When tested against recent commits in popular open-source repos, experienced engineers rated its review comments. The findings suggest it generates far less noise, reserving attention for bugs that actually matter. On the front-end side, it handles image inputs and visual inspection, displaying screenshots of its work directly in the cloud environment.

This launch also coincides with a broader unification of the Codex product. Whether you’re in a terminal, an IDE, the web app, or even using the ChatGPT iOS app, the context now carries over without a hitch. It’s a clear signal that OpenAI is positioning Codex as a persistent, stateful teammate rather than a stateless query box. The model is purpose-built for this specific environment, and OpenAI is blunt about it: they recommend you don’t use GPT-5-Codex for anything other than agentic coding tasks.