4 frontier models built their own chess boards and all lost to Claude Opus 5.
Really Interesting experiments by @thehypedotnews, a 24/7 AI news in a really nice radio format.
In this experiment, I find DeepSeek V4 Flash's performance really interesting.
It used 27.4 mn tokens, needed 5 attempts to build working stands, and still completed both tasks for $0.557.
So kind of changes how agent efficiency should be measured. A model can reason inefficiently at the token level and still remain economically useful if inference is cheap enough to make retries almost free.
qwen 3.8 max vs deepseek v4 flash 0731 vs kimi k3 vs gpt 5.6 sol - on rubik's cube and chess four frontier models built a rubik's cube stand and solved it, then...