Tag
GPT-6 Astra achieves a massive leap on ZeroBench, an extremely difficult vision benchmark, surpassing the human baseline across all three metrics.
The article describes testing the Qwen3.8-27B AI model with different quantizations and settings to recreate images as SVG, aiming to develop a benchmark resistant to benchmaxxing. Preliminary results indicate that high reasoning effort and specific cache configurations optimize performance.