Build More with GPT-5.1-Codex-Max

A new frontier agentic coding model, GPT-5.1-Codex-Max, has been unveiled. Optimized for agent-centric work such as software engineering, mathematics, and research, this model is faster, more intelligent, and more token-efficient than previous models.

Through its core technology, Compaction, it can consistently handle massive contexts spanning millions of tokens. This enables project-scale refactoring, deep debugging, and agentic loop tasks that run for hours at a time. According to internal evaluations, the model demonstrated the ability to work independently for up to 24 hours or more, autonomously fixing test failures and completing final deliverables.

Meaningful improvements in cost and efficiency were also confirmed.

• Token efficiency: On SWE-bench Verified, the model achieved higher performance while using 30% fewer thinking tokens compared to the previous model at the same reasoning level.
• Reasoning option: A new Extra High (xhigh) reasoning mode has been introduced, which thinks for longer periods to produce precise answers for tasks that are not latency-sensitive.

GPT-5.1-Codex-Max is now available immediately in Codex's CLI, IDE extension, cloud, and code review environments, with API access to be supported soon. It has also been designed to work smoothly on Windows environments.