DeepSeek V4.1 Flash nears GPT-6 Astra on everyday design benchmarks

SciTech Era @SciTecheraOriginally posted 2026.09.16

Open original on X

Open-source competition is getting crazy' 👀 But it's hard to believe some of these benchmark results. Could distillation or benchmark-specific optimization be playing a role? As per the chart, DeepSeek V4.1 Flash is still behind GPT-6 Astra in quality, but it's surprisingly close on everyday design benchmarks. It also reportedly costs 100× less to run and scores about the same on everyday design and other benchmarks. That's some crazy price-performance!

This post quotes another post: “DeepSeek V4.1 Flash said to cost 1% of GPT-6 Astra to train”