Strange that the system card carefully seems to avoid any benchmark where you can also find scores for GLM, Qwen. There's barely any overlap with GPT 5.6 benchmarks. Just these:

    Model              HLE w/tools   GDPval-AA v2 
    Claude Fable 5.1   65.0          1853
    GPT-5.6 Sol        64.5          ~1711-1730
    GLM-5.3            62.5          1769
    DeepSeek V4 Pro    60.0          1590
    Kimi K3            59.8          1682
    Qwen3.8-Max        56.2          1739