Comment by stingraycharles

“no harnass at all” might be an issue, though, as these types of benchmarks are often gamified and then models perform great on them without actually being better models.