...releasing a new model that’s marginally if at all better than the original...

This isn't what we see in benchmarks.