Citation needed. Not a microbenchmark, or a single example of some software. Actual sustained mixed workload usage of real life applications. Especially realtime-sensitive stuff like DAWs (where you have the added risk that calling into the JIT in the middle of a realtime thread can completely screw you over; keeping realtime-safe code realtime-safe under dynamic or even static binary translation is a whole extra can of worms).
Sustained benchmarks await production hardware. But it will be surprising if Rosetta2 translated apps run slower. Not only will system calls be native, but common operations like retain/release are 2x faster under Rosetta3.
That's a microbenchmark. There are a myriad reasons why one specific thing might be faster under a new CPU even under emulation. That doesn't mean other things won't be much slower.
