OpenAI launches GPT-5.1-Codex-Max with performance gains

OpenAI has launched GPT-5.1-Codex-Max, which is a new coding model and an upgrade to the predecessor. The company says the new model is built for long, complex engineering tasks, where an AI needs to plan, reason, and revise code across extended sessions without losing track of earlier steps.

One of the standout features of the GPT-5.1-Codex-Max is the new compaction system, which allows the model to keep massive tasks coherent, even when they run into the millions of tokens. This means developers can hand off bigger refactors, multi-file changes, and long debugging cycles without forcing a reset.

OpenAI further notes that the new model model is designed to stay on task for much longer, providing cleaner transitions between steps and better continuity.

OpenAI trained Codex-Max using real software engineering workflows, including pull requests, code reviews, and full-stack development. This makes it more effective in practical scenarios, not just benchmark tests. It also uses around 30% fewer thinking tokens than the previous GPT-5.1 Codex while still performing better.

Speaking of performance, the new model scores 77.9% compared to 73.7% with its predecessor on the SWE-Bench Verified benchmark. Not to forget, the model also supports multi-hour agent loops, which help AI agents maintain context and progress through extended work sessions.

OpenAI hasn’t compromised with safety either, as Codex-Max operates inside a restricted sandbox that limits file activity and blocks network access unless a user explicitly approves it.

Codex-Max is available today through the Codex app, CLI tools, cloud dashboard, code review system, and IDE extensions. OpenAI says API access will release soon, with Codex-Max will become the new default model across all Codex surfaces.