GPT-6 Luna 基准测试低推理档跑出假完成

Morgan @morganlinton原推发布于 2026.09.27

在 X 上打开原推

Waking up to see how my GPT-6 Luna benchmark is going with @VulcanBench 🖖 Now that Opus 5.5 is done, GPT 6 Luna and Sol are next. So far Low, Medium, and High have run, Extra High and Max still need to run. But I thought I'd share something interesting from the first three effort levels. What seems to be happening, at lower effort levels with Luna, is, it does't take the time it needs to solve the tasks correctly. Without more thinking depth, Luna just says, "jobs done!" but it's not, and it does it in like three minutes, when it has three hours to actually do it right. More details below, will share the full results once all effort levels are completed.