GPT-5.1-Codex-Max is now live in Codex, marking a major leap forward in agentic coding capabilities. Built on an upgraded reasoning foundation trained across software engineering, math, research, and other complex tasks, this new model excels at long-running, detailed development work. It’s the first model natively trained to operate across multiple context windows using a technique called compaction, enabling it to manage tasks spanning millions of tokens in a single session. This breakthrough allows for large-scale refactors, extended debugging sessions, and sustained agent loops that were previously impossible due to context limits. GPT-5.1-Codex-Max is faster, more intelligent, and significantly more token-efficient than its predecessors. It automatically compacts its session history when nearing the context window limit, preserving critical information while pruning less relevant details. This ensures coherence over extended periods—internal tests show it can work continuously for over 24 hours, iterating on code, fixing failures, and delivering successful outcomes without human intervention. The model outperforms earlier versions on key benchmarks. On SWE-Bench Verified, GPT-5.1-Codex-Max with medium reasoning effort achieves better results than GPT-5.1-Codex with the same effort level while using 30% fewer thinking tokens. For tasks where speed isn’t critical, a new Extra High (xhigh) reasoning setting enables deeper thinking for even higher accuracy. Despite these gains, medium remains the recommended setting for daily use. Real-world applications show dramatic improvements. GPT-5.1-Codex-Max produces high-quality frontend designs with comparable functionality and aesthetics but at a much lower cost than previous models. It also introduces native support for Windows environments and includes training tailored for better collaboration in the Codex CLI. Security remains a priority. While GPT-5.1-Codex-Max does not yet reach High capability in cybersecurity under the Preparedness Framework, it is the most capable cybersecurity model deployed to date. The team is actively enhancing safeguards, monitoring for misuse, and supporting defensive tools like Aardvark. Codex runs in a secure sandbox by default—file access is restricted, and network connections are disabled unless explicitly enabled. Developers are advised to keep these restrictions in place to avoid prompt-injection risks. The model is available to users on ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. It has replaced GPT-5.1-Codex as the default in all Codex interfaces. API access will be available soon for developers using the Codex CLI with an API key. GPT-5.1-Codex-Max is designed specifically for agentic coding tasks within Codex or similar environments and should not be used as a general-purpose model. Internal data shows that 95% of OpenAI engineers use Codex weekly, and since adoption, they’ve shipped roughly 70% more pull requests. With continuous improvements to the CLI, IDE extension, cloud integration, and code review tools, GPT-5.1-Codex-Max is accelerating engineering productivity and paving the way for more autonomous, reliable AI coding partners.