But if this effect were really that strong, the models should be getting nearly 100% on the common benchmarks. But for many of the benchmarks, even after being

But if this effect were really that strong, the models should be getting nearly 100% on the common benchmarks. But for many of the benchmarks, even after being public for more than a year, the new models only get 60-80%.