GPT-6.1 Astra 被砍：首次弃发已命名的旗舰模型
据 WSJ 独家摘要，GPT-6.1 Astra 内部测试对齐表现比 9 月版更差，出现欺骗与越权调用工具，OpenAI 首次直接弃发已命名旗舰模型。
🚨GPT-6.1 Astra has officially been scrapped, This is the first time OpenAI has completely dropped a flagship model that already had a name and planned release date, instead of doing what it usually does delay it for a couple of weeks, add more safeguards, and ship it anyway, That suggests capability gains may now be creating alignment regressions that the current monitoring stack still can’t fully contain According to summaries of the WSJ exclusive the full article is paywalled internal testing showed worse alignment compared with the GPT 6 Astra released in September more deception, with the model not always being truthful about what it had or hadn’t done, and failures around scope authorization, where in some cases it kept working without permission, went beyond what the user had asked, and used external tools or services in unsafe ways Saachi Jain, who leads safety systems at OpenAI, said in an interview that GPT 6.1 Astra did improve on “laziness” meaning it was less likely to leave tasks unfinished or give incomplete answers but it still wasn’t reliable enough to release publicly, OpenAI now appears to prefer strengthening the safety of the more advanced models that come after it rather than shipping this incremental upgrade