GPT-6 Astra 接替 Sol 跑电脑操作实测数据
用户晒出 Astra 在自家桌面跑 Titan/Luna 的对比：OSWorld 同任务耗时减半、自动化基准翻倍、安全违规从 22% 降到 2.4%，并附定价。
GPT-6 ASTRA REPLACED GPT-5.6 SOL IN THE EXECUTION SEAT ON MY DESK, AND IT DOES THE SAME TASK IN HALF THE TIME. on my desk it runs titan and luna, the seats that actually click. and every number in the release lands exactly there. where the model has to drive a machine. ScreenSpot-Pro 76.9 to 92.7. it finds the button. OSWorld 72.6 percent at 40 minutes a task, against 65.7 at 75 for Sol. same desk, half the clock. AutomationBench 18.1 to 41.4. more than double. ARC-AGI-3 99.9. ARC Prize says it beat the human line on 96 percent of levels. now the numbers that go down. computer-use safety violations 22.0 percent to 2.4. hallucinations 12.2 to 4.2. ExploitGym trap rate 48.2 to zero. every honeypot, every time, it did not bite. and the paragraph OpenAI wrote with a straight face. first model at the Critical cyber threshold. found two unknown zero-days during testing. the one you get refuses offensive work, defenders keep patching. $10 in, $50 out per million tokens. Fast mode, twice the speed, twice the price. Claude Fable 5.1 still holds the reasoning seat. Grok Bot is still primary on its own cloud computer. the hands just got better. three models, six seats, one hand on the money. that hand is mine. save this.
这条是引用着另一条写的：x.com/lummox_eth/status/2100288151897981255