OpenAI releases GPT-5.1-Codex-Max

Background

Coding models have moved from completing a line of code towards what is often called agentic coding: a person states a goal, and the model works through the steps itself — reading files, editing code, running commands, checking results — instead of answering a single request in one pass. Once tasks stretch out, two things come under pressure at the same time: how long a model can keep the goal in view, and how many tokens (the units of text a model reads and writes) each step costs.

What happened

On 19 November 2025, OpenAI released GPT-5.1-Codex-Max, an agentic coding model ^[1]. The same day, the company released GPT-5.1 Pro, which replaced GPT-5 Pro ^[1].

GPT-5.1-Codex-Max became one of the model options inside OpenAI's Codex agent, the environment in which the company's coding models carry out tasks ^[1].

Key details

OpenAI said the model was built on an update to its foundational reasoning model, and that it was faster, more intelligent and more token-efficient at every stage of the development cycle ^[3]. Token efficiency matters for agentic work because a long task can consume a great deal of text: if each step costs fewer tokens, the same budget covers more of the job.

The model was trained on real-world software engineering tasks, including pull request creation, code review, frontend coding and code Q&A ^[3]. These are routine activities of a working software team rather than isolated puzzle problems. A pull request is a proposed change a developer submits for others to read and comment on; code review is that reading and commenting; frontend coding is the part of a product that users see and operate; and code Q&A means answering questions about an existing codebase.

VentureBeat reported that, in internal testing, the model had worked on a task for 24 hours ^[2]. Sustaining work over that span is a different proposition from producing a quick answer. The longer a task runs, the more the model has to hold the goal steady across many intermediate steps, and the more each of those steps has to remain consistent with what came before it.

OpenAI also drew a boundary around the Codex family: it recommended those models only for agentic coding tasks in Codex or Codex-like environments, unlike the general-purpose GPT-5.1 ^[3]. The distinction is about where a model is expected to be used. A general-purpose model is asked to handle a wide range of requests; a coding model tuned on engineering tasks is being pointed at a narrower setting, with the tools and files of a coding environment around it.

Why it matters

The release put two lines of work side by side on the same day. GPT-5.1-Codex-Max went into the Codex agent as a model option for coding work, while GPT-5.1 Pro took the place of GPT-5 Pro ^[1]. The task list the coding model was trained on — pull requests, review, frontend work, questions about code — describes stages of an ordinary engineering workflow rather than a single benchmark exercise ^[3].

Another signal is the time scale. A model reported to have spent 24 hours on one task in internal testing is being measured on how far it can carry a piece of work, not only on how good a single answer is ^[2]. OpenAI's recommendation that the Codex family be used in Codex or Codex-like environments makes the same point from the other direction: the company frames these models as tools for a particular kind of setting rather than as drop-in replacements for a general assistant ^[3].