Yes, that’s what he claimed here, too. But he is known to have changed his mind over the years and gaslight others in forums.
That’s why I was very specific, I was referring to this document from the early 2000s authored solely by Coggan (taken from the top of p. 5):
In that document he also discusses other ways to estimate power at MLSS, including critical power.
Going by what he wrote (including one of his posts on that subject here), he did agree that maximal 20-minute power minus 5 % is a viable way to estimate MLSS power — on average.
Independently of what Coggan thought or thinks, for many, many years a 20-minute test was the de factor way to determine FTP in practice. Just like with any test, with time athletes gained experience and then simply corrected the numbers accordingly to get an estimate for their power at MLSS from a test result.
Importantly, TR’s AI FTP no longer wants to approximate power at MLSS, but be able to offer you the right workout. That’s subtly different.
How long you can do power at MLSS depends on you and your training. That’s always been an issue with any FTP test, any simple computation based on statistical averages will fail people as most of us are not average in that respect.
Machine Learning changes things here, if done properly, it goes beyond simple statistical relationships and instead bases its predictions off of the past performance of an athlete. That’s why I wrote AI FTP has the potential of being a much better way to gauge FTP than any test protocol.