OpenAI’s new GPT-5.1-Codex-Max — everything you need to know about the coding model built to work for long hours

OpenAI has introduced GPT-5.1-Codex-Max, a powerful new coding model designed to handle long, complex software development tasks. Here’s a clear look at what it is and why it matters.

---

OpenAI this week released GPT-5.1-Codex-Max, an upgraded “agentic” coding model built to take on long-running, detailed software work without slowing down.

The company describes the model as faster, smarter, and far more efficient with tokens—essentially meaning it thinks better, works quicker, and costs less to run. It’s now available across all Codex products.

The timing is interesting—this release comes just after Google announced its developer-focused AI system, Antigravity. So yes, the two tech giants seem to be gearing up for a serious rivalry in the future of AI-powered software development.

What is GPT-5.1-Codex-Max?

Codex-Max is OpenAI’s most advanced coding model yet, built on the GPT-5.1 architecture. Instead of being trained only on generic code examples, it learned from real software engineering tasks—writing pull requests, reviewing code, building websites, and even answering tough technical questions.

Because of this specialized training, it outperforms every previous OpenAI coding model on high-level coding evaluations. And for everyday use, it brings something many developers have been asking for: it now works smoothly on Windows. OpenAI also trained it to behave more like a helpful teammate when using the Codex command-line tool.

One of the most impressive parts? Codex-Max can work independently for extremely long stretches. During internal testing, the model kept improving its own code, fixing bugs, and completing tasks even after running for more than 24 hours nonstop. That’s the kind of stamina human developers can only dream of.

OpenAI CEO Sam Altman praised the team behind the model, calling them “beasts” for the speed and quality of their progress.

It has been amazing to watch the progress of the Codex team; they are beasts.

The product/model is already so good and will get much better; I believe they will create the best and most important product in the space, and enable so much downstream work.

— Sam Altman (@sama) November 23, 2025

“The product/model is already so good and will get much better,” he wrote on X. “I believe they will create the best and most important product in the space, and enable so much downstream work.”

Who can use this new model?

GPT-5.1-Codex-Max is available to users on ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. Developers using the Codex CLI with an API key will get access once API support rolls out.

From now on, GPT-5.1-Codex-Max will replace GPT-5.1-Codex as the default model in all Codex interfaces.

OpenAI also shared that 95% of its internal engineering team uses Codex every week—and since adopting it, engineers are shipping about 70% more pull requests. Clearly, the tool is already reshaping their workflow.

Accuracy and efficiency

OpenAI says GPT-5.1-Codex-Max is a significant upgrade. In the SWE-Lancer coding test, it answered 79.9% of questions correctly—up from 66.3% with the previous GPT-5.1-Codex model.

In another benchmark (SWE-bench Verified), it solved more tasks with higher accuracy while using about 30% fewer thinking tokens. In plain language: it gets more work done while thinking less, making it cheaper and faster for developers.

In one internal example, Codex-Max built a full browser-based CartPole reinforcement learning sandbox using only 27,000 thinking tokens, compared to 37,000 used by the earlier model.

To give developers even more control, OpenAI is also adding a new “extra-high reasoning” mode for tasks where speed doesn’t matter. This lets the model spend more time thinking deeply before giving an answer—like a programmer who stops, breathes, and really focuses on solving a tough bug.