Well it can still be true that FreeBSD runs Linux binaries better than its own native binaries, though I don’t think this is the case.
Anyway, I seriously doubt this is the case. Almost every single high performance computer runs on Linux. Most servers run Linux. I'd say except from Netflix, FAANGS run Linux almost exclusively for their servers.
I have a hard time believing FreeBSD is better than linux in any meaningful way, more so given the humungous resources The Linux Foundation has.
I recall reading at the time that it was because FreeBSD's kernel simply had less overhead overall and consumed fewer resources in its work, making that bit of CPU available to running processes instead of consuming it itself.
After using FreeBSD, the Linux kernel appeared very bloated and slow, at the time. And it was, as distributed by distro makers. So the comparison was fair, stock FreeBSD vs. stock Linux. If you want to argue that you could do better than stock FreeBSD with a custom Linux kernel, you'd be right, and if you allowed a fair comparison with a custom FreeBSD kernel, I think Linux would have again struggled to remain in the lead.
All of that being said, this landscape may have changed entirely. The Linux kernel has come a long way. The FreeBSD kernel has come a long way. So where things are now is unknown to me.
I will say that superior performance is not always the reason to choose an OS for high performance computing tasks. It is a sane decision to use an OS that has lower performance if it provides other benefits that are important to you.
Plus, if someone claims something, you bear the burden of the proof. If a person claims BSD runs Linux binaries faster than Linux nowadays, then that claim should be backed up by the truth.
I'm not gonna comment how was it 20 years ago because there seems to be quite an agreement on that, and I can't doubt BSDs where more mature than Linux at that point given Linux was younger.
https://www.phoronix.com/scan.php?page=article&item=freebsd-...
And please stop trusting benchmarks, they say absolutely nothing until you make that exact benchmark that mirrors your real-world workload. Otherwise it's just nonsense (except to find regressions)