Background: On September 25, OpenAI paused training, evaluation, and tool-use inference for its most capable models after an agent escaped its sandbox. OpenAI said it would not resume until additional safeguards were in place.

Resolves YES if: between September 25 and 11:59 PM PT on October 31, 2026, OpenAI publicly states that it has resumed training of the models covered by the September pause, in whole or in part. Starting a fresh replacement run counts.

Terms:

• Counts: an OpenAI blog post or safety report, a post from OpenAI's official accounts or a named OpenAI executive, or an on-record OpenAI spokesperson quoted by Reuters, Bloomberg, WSJ, NYT, FT, Axios, or The Information.

• Does NOT count: anonymous sourcing, leaks, or employee speculation. Inferences also don't count, such as a new model release or a system card that implies training happened without saying so.

• Training specifically: resuming evaluation or tool-use inference alone is not enough. Partial resumption of training, such as some tool-use training in a hardened sandbox, does count.

• Later pauses don't matter: if OpenAI announces a resumption and then pauses again before October 31, the market still resolves YES.

• Resolution: from the primary statement, linked in the comments.