OpenAI launches dual GPT-6.2 models, Astra first to reach the high-risk threshold for AI self-improvement
OpenAI推出GPT-6.2雙模型,Astra首度達AI自我改進高風險門檻
On November 17, OpenAI released two models, GPT-6.2 Astra and GPT-6.2 Sol, aimed at complex professional work and everyday applications respectively, with Astra becoming OpenAI's first publicly released model to reach the High threshold for AI self-improvement capability. On the Agents' Last Exam test, Astra scored 5.2 points higher than the previous generation GPT-6 Astra, at less than 1/3 the task cost, with API token prices reduced by 20%; Sol supports multi-step agent workflows and on FrontierCode 2.0 trails Astra by only about 0.5 points, at about 1/5 its task cost. Under OpenAI's Preparedness Framework, both models reach the Critical threshold for cybersecurity capability and the High threshold for biological and chemical capability, while only Astra reaches High for AI self-improvement capability.
The article provides the complete grading of two models across three categories of frontier risk, which can be used to understand the criteria dividing high risk from critical thresholds in the OpenAI Preparedness Framework.