A post from @teortaxesTex is circulating with the claim that Grok 4.5 tied Sol on KernelBench-Hard without reward hacking, alongside a benchmark image readers can compare against KernelBench-Hard results. So far, the public evidence is still narrow, so the safest read is the attributed claim itself, plus the limited source context available in the post and linked benchmark page. That makes this more of a notable public claim than a settled result.