According to ChainCatcher, OpenAI team member Tibo said GPT-6.2 Astra is adept at difficult tasks and appears to lead globally on the hardest tasks. He said the team is working to improve the model’s communication style in response to feedback. Tibo also cited up to 14% improvements in power, performance and area on flagship GPU IP through OpenAI’s work with AMD, and said the model helped optimize OpenAI’s first custom inference chip, Jalapeño, with improvements verified on the chip. OpenAI evaluations showed GPT-6.2 Astra had the highest average score across two open-source RISC-V processor core tests, using a comparable budget of about 112,000 tokens per run.