The Information reported, as relayed by The Decoder on 14 October, that OpenAI staff expect one or two new models by year's end. It is unclear whether the major upgraded model in development is one of them. A person familiar said that flagship uses looped transformers, which send each token through the same layers several times before the next word. OpenAI already uses this in Astra, its top model.
The Information says the flagship may be well off. It needs alignment work, so its goals match what researchers and customers expect, plus stronger testing infrastructure. A person familiar with the figures said OpenAI shifted roughly a tenth of its total compute to alignment and safety after the Hugging Face incident, which the report does not explain. The Information puts that at hundreds of megawatts. The earlier safety share and the current split are unclear.
Looping cuts costs and lifts math and coding results, but it moves some reasoning out of readable text, so the saving and the audit problem share one cause. OpenAI's chief scientist has said monitoring is growing fragile. Buyers can ask vendors whether loop counts are capped, as OpenAI did for Astra, and what reasoning trace they can audit.