OpenAI releases GPT-5.1-Codex-Max for long-running coding tasks
A 'compaction' technique lets the model summarise and clear its own context automatically, and OpenAI reported sessions running over 24 hours in internal testing.
- Models & capabilities
- Minor
OpenAI released GPT-5.1-Codex-Max, a coding-focused model built to work through a technique it calls compaction: when a session approaches its context-window limit, the model automatically summarises and clears earlier context rather than stopping, letting it continue working across multiple windows. OpenAI said internal testing showed sessions running for more than 24 hours.
On coding benchmarks run with compaction enabled at the model’s highest reasoning effort, OpenAI reported SWE-Bench Verified rising from 73.7% to 77.9%, SWE-Lancer’s independent-contractor coding tasks from 66.3% to 79.9%, and Terminal-Bench 2.0 from 52.8% to 58.1%, relative to GPT-5.1-Codex, the model released a week earlier. OpenAI also said the new model reached comparable results on SWE-Bench Verified using roughly 30% fewer reasoning tokens, framing it as a cost and speed improvement as well as a capability one. As with other model-release benchmarks, the figures were run by OpenAI itself rather than reproduced independently.
The model shipped in the Codex CLI, IDE extension, cloud environment and code-review tooling, extending a pattern of frontier labs releasing separate coding-specialised variants of general models — following GPT-5-Codex two months earlier — aimed at the agentic, long-running software tasks that had become one of the more commercially visible uses of large language models.