OpenAI has released GPT-5.1-Codex-Max, an agentic coding model built to run uninterrupted for more than 24 hours on a single development task.
The company says it is the first system able to work across multiple context windows through a technique called compaction, which allows it to stay coherent across millions of tokens. Internal tests showed Codex-Max fixing test failures and improving code unsupervised for more than a day.
The model is now available to ChatGPT Plus, Pro, Business, Edu and Enterprise users and will replace the previous Codex model as the default. API access via Codex CLI will follow “soon”, OpenAI said.
Engineers inside OpenAI already generate roughly 70% more pull requests since adopting Codex, the firm claims. On the SWE-Lancer benchmark Codex-Max scored 79.9%, up from 66.3%, while using 30% fewer reasoning tokens. A browser-based CartPole reinforcement-learning sandbox that once required 37,000 tokens now needs 27,000.
An extra-high reasoning mode lets the model spend additional time on non-urgent tasks. Compaction also enables complex refactors and multi-hour debugging loops that previously failed due to context limits. Codex-Max is the first model in the series with Windows support and includes tasks aimed at improving collaboration through the CLI.
OpenAI rates the model below “High” on its internal cybersecurity scale, but says it is the most capable security-oriented version deployed so far. It runs inside a restricted sandbox with no network access unless explicitly enabled. The company recommends treating Codex “as an additional reviewer, not a replacement for human review”.