GPT-6 Astra High tops PPTBench at 77.34 but still misses ~19% of tasks

Rose Carter @Bisma7465Originally posted 2026.09.29

Open original on X

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.

This post quotes another post: x.com/einsiaai/status/2104941770433814889