OpenAI describes its coming model Astra as stronger at coding and at operating applications on a computer. The gains come partly from a new training technique. That technique also means Astra shows less of its internal reasoning. Models built this way are therefore harder to monitor for unsafe behaviour. Researchers have relied on readable reasoning traces to check how a model reached an answer. Astra has not been released and OpenAI has not given a date. The account comes from reporting by The Information, which cites OpenAI on the capability gains.