A Trezor acts as firewall between "the wallet that can be easily stolen from an USB stick" and your host.
Given that a Trezor and a pure offline machine are pretty similar, I tend to disagree here. Pro argument for Trezor: a pure offline machine is still a much more complex machine than a tiny single purpose device, even though, in reality, it's actually not a single purpose device, but a mini version of the complex machine, further restricted by some software as per default.
Assuming both Trezor as well as offline machine are never ever exposed, then they are even. Exposing one of them adds risk, that is without any doubt. A Trezor is likely more secure in an exposed context.
But the initial question was: is an exposed Trezor more secure than annever exposed offline wallet, right?
Well, just keep the offline machine isolated. Create your keys offline, sign your transactions offline. There is no need to connect the offline machine to an online machine or any other machine, ever. Moving data between devices via USB is by no means required, assuming you move data by hand or some other air gapped mechanism. Start here, if you want to use Bitcoin Core, but I'm pretty sure Electrum and especially Armory provide a way to handle offline transactions as well. Think of it this way: where you act as connection between the offline machine and the online machine by moving data around by hand, the Trezor basically does the same, whereby the firmware of the device fulfills your role of moving only the data that is considered as accepted. I think the key difference and the lack of comparsion in the whole discussion is rather based on the assumption that a Trezor is usually connected more frequently, because that it's purpose, while a pure offline wallet is usually used very rarely.
Fully agree. Trezor is a great device and adds a nice layer of protection.
Nevertheless and by the way, it sucks to see some of you guys actually flame gweedo. You may argue about the way his problem was presented, but this doesn't magically solve any problem, which is, without a doubt, exisiting. That said, one of my first experiments with my Trezor was the attempt to feed it with data it probably doesn't understand, namely m-of-n multisig transactions. Guess what: the result was a never ending loop of errors which resulted in a complete browser freeze which I was only able to tame after unplugging the device. That being said, if this wasn't done intentionally and if real coins were involved - which I still could not access (for whatever reason that might be) several days later.. oh well.
My experience with the support was actually fine, with a response in probably less than 6 hours, assuring me that this problem is "known and taken care of by the devs".