OpenAI today released GPT‑5.1-Codex-Max, a new generation Agentic coding model designed for long-running tasks. Compared with previous models, GPT‑5.1-Codex-Max uses "compaction" technology, which can work across multiple context windows and can even reliably process millions of tokens in a single task. OpenAI said that the model not only improves performance, but also achieves faster and more efficient token utilization.
It is understood that the GPT‑5.1-Codex-Max training process covers real-world software engineering tasks, such as PR creation, code review, front-end development and question and answer, etc., and has outperformed previous models in many cutting-edge programming evaluations. For example, the model achieved a score of 77.9% on SWE-Bench Verified (500 samples), 79.9% on the SWE-Lancer IC SWE evaluation, and 58.1% on the TerminalBench 2.0 evaluation, all higher than the previous performance of GPT-5.1-Codex.
In addition to supporting Unix platforms, GPT‑5.1-Codex-Max is specifically trained for Windows environments. In complex reconstruction and long-running agent loops, most coding models on the market are limited by the context window and are difficult to work continuously. GPT‑5.1-Codex-Max, on the other hand, can run autonomously for hours or even tens of hours by automatically compressing session content when it approaches the window limit. According to OpenAI internal test data, the model can run continuously for more than 24 hours.
In addition, thanks to improved reasoning capabilities, GPT‑5.1-Codex-Max uses 30% less thinking tokens than GPT-5.1-Codex when completing the same task on SWE-Bench Verified. Through the "Extra High (xhigh)" reasoning mode, the model can engage in deeper thinking in complex tasks.
Currently, GPT‑5.1-Codex-Max has been launched in Codex CLI, IDE extension, cloud and code review products, supporting ChatGPT Plus, Pro, Business, Edu and Enterprise premium subscriber users. At the same time, OpenAI will also introduce this model into the API in the near future and replace it with the default model in Codex.