Fresh benchmark drops Gemini 3.8 Flash and Muse Spark 1.3 from ~90% to 19% and 33%

Dan @DanDr1sOriginally posted 2026.09.08

Open original on X

Google and Meta are benchmaxxing hard. Gemini 3.8 Flash and Muse Spark 1.3 look almost as good as GPT-6 Astra and Fable 5.1. Then a fresh benchmark drops: Gemini: 89.4% → 19.1% Muse: 88.8% → 33.3% Meanwhile GPT-6 and Fable still score ~56–58%. This is what happens when labs optimize around public benchmarks instead of building models that actually generalize. Benchmark scores are becoming harder and harder to trust.

This post quotes another post: x.com/semianalysis_/status/2097112794717958498