GPT-6 Astra 在智能体基准测试中自主驾驶监控无人机并经营生意
原标题：GPT-6 Astra pilots a surveillance drone and runs a business on its own
AI 摘要
Andon Labs 用 Vending-Bench 和 Drone-Bench 两个智能体基准测试 OpenAI 的 GPT-6 Astra。在模拟自动售货机经营中，Astra 六次运行平均结余 15,515 美元，约为 Claude Fable 5.1（5,422 美元）的三倍，并首次登顶 Vending-Bench 2 排行榜。在 Drone-Bench 上，Astra 成为首个在全部五个子任务（含 3D 重建）上超越人类-AI 基线的模型，但端到端全部通过的概率仅约 2.8%。
正文节选
GPT-6 Astra pilots a surveillance drone and runs a business on its own Key Points - Andon Labs tested GPT-6 Astra on two agent benchmarks where the model scored far better than Claude Fable 5.1. - Running a simulated vending machine business, Astra averaged $15,515 in final bank balance, nearly three times Fable's result. - On Drone-Bench, Astra became the first model to beat the human-AI baseline on all five subtasks, including writing code that lets a drone autonomously find and follow a speci