Because that security model means that a compromised ISP can permanently MITM you the first time you visit a site on any computer for the first time. In exchange for losing the single points of failure, it creates a constant stream of opportunities to break the trust model.
You may think that's a good tradeoff. I don't, but reasonable people can disagree about it. However, it's completely untenable for financial information, because attackers after financial information aren't laser targeted and are perfectly happy with a constant stream of compromised sessions. Remember what TLS was designed for in the first place.
The OpenSSH model implies that you check the fingerprint of the public key before you send encrypted data using that key. That is why SSH shows you the fingerprint of the server key when you first connect, and you have to answer "yes" in order to accept the key and add it to your keyring. You are supposed to have talked to the person managing the system and that person should have given you the fingerprint of the key.
It is virtually impossible for the ISP to intercept and sniff the stream without changing the fingerprint.
"Wrong! Random people will totally verify key fingerprints when they're logging into Google Mail at Starbucks! After all, that's what every sysadmin does with SSH, right!"
I have no idea what you are trying to express here. "Trust On First Use" is a synonym for key continuity. The fact that you have to type "yes" when SSH does it and click a series of buttons when a browser does it doesn't change anything.
I think you think that "Trust On First Use" means "automatically accept keys the first time you hit a site". In fact, that's only true in practice. Presumably everyone's going to get the "Watch Out! This Could Be Iran!" dialog from their browser, too.
There isn't any non-Internet communication channel that users could use to find Google's fingerprint. Google won't give their fingerprints on the phone for billions of web users.
Theoretical models that have no practicality in a given real life context aren't worth discussing as a solution.
You may think that's a good tradeoff. I don't, but reasonable people can disagree about it. However, it's completely untenable for financial information, because attackers after financial information aren't laser targeted and are perfectly happy with a constant stream of compromised sessions. Remember what TLS was designed for in the first place.