OpenAI GPT-6.2 Astra criticized by testers as the worst on its first day of release; internal engineers reportedly have abandoned it
OpenAI翻车!GPT-6.2 Astra被批「史上最差」,内部员工连夜弃用
OpenAI released GPT-6.2 Astra and it drew concentrated negative reviews from early testers on the first day; Theo Browne called it the worst model release in OpenAI's history. According to multiple sources, OpenAI's internal engineers have abandoned Astra in their daily work and switched to GPT-6.2 Sol Ultrafast. Browne ran Astra continuously for 7 days on the ts-rust project, and the number of improvements obtained was 0; it also slowed the compiler by 56% because it fixated on the prompt's requirement to reduce size, and tests by the Every team also showed tasks stopping prematurely and degraded writing ability. The article also says Astra is the first model to achieve a 100% score on ARC-AGI-3, using 64.1% fewer operations than the human baseline when running 25 public games at the highest reasoning level, at less than one-third the cost of GPT-6 Astra. The article argues that over-alignment leads it to choose to do less or fall into loops in order to avoid major mistakes.