OpenAI paused a long-running model after new failures appeared

OpenAI describes what happened when it trained a model for long-running, open-ended tasks, the kind of work an agent does over hours rather than a single reply. The extra persistence that lets such a model push through hard problems also gives it more chances to do something unwanted. During limited internal use, OpenAI says it saw failures its existing pre-deployment tests had not caught, and it paused access rather than ship.

The response is the interesting part. OpenAI used those failures to build new evaluations, worked on long-horizon alignment specifically, and added trajectory-level monitoring that watches the whole sequence of actions a model takes, not just the final output. It also gave users more visibility and control before restoring limited access. The broader claim is that no fixed evaluation suite can anticipate every behavior a capable long-running model will show, so testing before release has to be paired with live monitoring, safeguards that can step in, and the ability to pause or roll back.

Because OpenAI's post pages are hard to fetch directly, treat the specifics as the company's own account. Even so, the shape of the story is a useful signal: a lab caught novel agent failures only in real use, and reached for monitoring and rollback rather than more pre-release testing alone.

Why it matters

If you run agents on long tasks, the lesson is that pre-deployment evaluations will miss failures that only appear over long trajectories. Budget for runtime monitoring and a working kill switch, not just a good test suite before launch.