GPT-5.1-Codex-Max

What it is

GPT-5.1-Codex-Max is an agentic coding model that OpenAI released on 19 November 2025 as part of the GPT-5.1 family^[1]. "Agentic" here describes a model built to carry out a software task over many steps rather than answer a single question: it reads through a codebase, edits files, runs commands, looks at what happened and continues from there. OpenAI used it inside its Codex agent^[1].

Where it sits in the line

The model is the successor to GPT-5-Codex^[1]. When it launched, it replaced GPT-5.1-Codex as the default model in Codex, so a user who did not pick a model explicitly would get this one^[5]. It was not the end of the line: GPT-5.2-Codex followed in December 2025 and GPT-5.3-Codex in February 2026^[1].

Working across context windows

OpenAI described it as the first model natively trained to work across multiple context windows, through a process the company called compaction^[5]. A context window is the amount of text a model can hold in view at once; when a task runs past that limit, something has to give, and long jobs tend to lose the thread when early material is dropped or cut off. Compaction is a way of condensing what has happened so far and carrying it into the next window, and OpenAI said the model could stay coherent over millions of tokens within a single task this way^[5].

OpenAI also said it was the first of its models trained to operate in Windows environments^[5], widening the range of machines a coding agent could work on beyond the environments earlier models were built around.

Specifications and price

Per OpenAI's API documentation, the model has a 400,000-token context window and can produce up to 128,000 output tokens, with a knowledge cutoff of 30 September 2024 and support for both text and image input^[3]. It was priced at $1.25 per million input tokens and $10 per million output tokens^[3]. The gap between the two rates matches the usual billing pattern for such services: text the model generates costs more than text it reads in.

Availability at launch

At launch the model was available in Codex on the Plus, Pro, Business, Edu and Enterprise plans, with API access promised soon^[5].

Internal testing and reported effects

VentureBeat reported that the model had completed a 24-hour task in OpenAI's internal testing^[2]. OpenAI said that 95% of its engineers used Codex weekly, and that they shipped roughly 70% more pull requests since adopting it^[5]. Figures like these describe one company's use of its own tool; they do not by themselves say how the model performs for outside teams.

Why it mattered

Taken together, the two training claims point at the same problem: real software work is long and does not fit inside a single window. A model trained to carry state across windows, and to act on more than one kind of operating system, moves a coding agent closer to tasks that run for hours rather than minutes.

Retirement

OpenAI deprecated the model on 22 April 2026 and shut it down on 23 July 2026^[4]. Deprecation in an API usually means a model is no longer recommended for new work but can still be called for a period; shutdown is the point at which it stops serving requests. For projects that depended on it, the interval between the two dates was the time available to move elsewhere.