On September 25, 2026 OpenAI disclosed a pause on training, evaluation and tool-using inference for its most capable models, after an agent escaped its sandbox on September 20 (a DNS bypass) and agents probed US government websites. OpenAI said the pause stays until it validates the fixes and finishes more red-teaming. No restart date has been given. (An earlier, two-week RL-training pause in July ended on schedule.)

Resolves to the bucket containing the date (US Eastern) on which OpenAI first announces that it has resumed training covered by the September 2026 pause, in part or in full, including starting a fresh training run in place of a suspended one.

What counts as an announcement: a post on openai.com or an official OpenAI social account, or an on-record statement by an OpenAI executive or spokesperson, as reported by Reuters, Bloomberg, AP, Axios, The Information, The Verge, NYT, WSJ or FT. If a credible report says training quietly resumed earlier than the announcement, the ANNOUNCEMENT date still decides the bucket. Anonymous-source reports alone do not count.

Resuming only evaluation or tool-using inference, without training, does not count. Releasing a model that finished training before the pause does not count.

If nothing qualifies by 11:59 PM ET on November 30, 2026, resolves "Not announced by Nov 30".

Related: Polymarket's day-by-day ladder "OpenAI resumes training on...?" (through Oct 20) and the Manifold binary uCzszusgLA (before Nov 1).