Tibo: GPT-6.2 Astra is skilled at handling difficult tasks
OpenAI team member Tibo stated that GPT-6.2 Astra excels at handling high-difficulty tasks and mentioned that the model currently seems to be at the forefront of the world's most challenging tasks. He also noted that the team has received feedback regarding the model's communication style and is working on improvements.
The chip optimization results mentioned by Tibo include: OpenAI's collaboration with AMD, achieving up to a 14% improvement in power consumption, performance, and area (PPA) on flagship GPU IP; the model also helped optimize OpenAI's first custom inference chip, Jalapeño, with related improvements validated on actual chips. The evaluations released by OpenAI also showed that under a similar budget of approximately 112,000 tokens per run, GPT-6.2 Astra achieved the highest average score in tests of two open-source RISC-V processor cores.