GPT-6 Astra High 登顶 PPTBench 但仍有短板

Rose Carter @Bisma7465原推发布于 2026.09.29

在 X 上打开原推

GPT-6 Astra High leads PPTBench at 77.34, yet still misses a gate on roughly 19% of tasks. The bigger takeaway is how clearly PPTBench exposes failure points: semantics first, followed by render/text and visual details. It’s a useful reminder that benchmarks matter most when they reveal where models break instead of making them look better than they are.

这条是引用着另一条写的：x.com/einsiaai/status/2104941770433814889