OpenAI announced the launch of GPT-5.1-Codex-Max, designed for complex software engineering projects, supporting context consistency generation for tens of thousands of lines of code. The new model introduces a "compaction" dynamic compression mechanism, which can automatically organize memory and retain key states during task execution, significantly reducing the risk of information loss in long sessions.
Performance Data
- The accuracy rate on SWE-bench increased from 73.7% to 77.9%, with independent contribution software engineer tasks jumping to 79.9%, and token consumption reduced by approximately 12%.
- The "illicit" content detection score increased from 0.860 to 0.920, but OpenAI noted that cybersecurity capabilities have not yet reached the "high capability" standard and still require manual review.
Codex-Max has replaced the previous version as the default model in the Codex series, available to developers and enterprises through ChatGPT Enterprise, API, and GitHub Copilot. The pricing remains at $5 per million input tokens and $15 per output token, with a 50% discount for bulk calls. OpenAI plans to launch the "Codex-Max-Enterprise" dedicated version in Q1 2026, supporting private deployment and custom code style rules.
