It generally will stop them. The phrase that applies is something like "you don't have to outrun the cheetah, just another gazelle". It's not about eliminating the spammers, but about making it expensive enough - in CPU, memory, bandwidth, time, human intervention, etc. - that it's not worth them attacking you.

These days a spammer could train a ConvNet to pass reCAPTCHA with > 90% accuracy and very little processing overhead if they really wanted to. The only reason it works is because the bar is "high enough" that it's cheaper to spam somewhere else that has a lower bar.

It doesn't stop them, but it raises the costs of it significantly. Instead of hundreds of requests a second, they're down to one request every 5 seconds, and they're having to run the computer at a full CPU load 100% of the time. It would be more profitable for them to just mine bitcoins at this point, meaning they wouldn't waste that CPU load on spam submissions.

You're mixing the two cases. If every page is limited then yes they have to work hard, but still get delay/machines visits per second. But a human will have to wait the full delay every time a page is visited. This is unacceptable for modern browsing.

If the delay is only once per, let's say a domain, then you don't do anything against the spammers, they only have to wait a full delay once.

Maybe Cloudflare could have a browser plugin that preemptively creates "tokens", or essentially just mines Cloudcoins that you then spend to bypass CAPTCHAs. That way you could make it much more expensive than 5 seconds of CPU and there'd be zero delay (or perhaps not even the Cloudflare splash page). The use case for needing to constantly bypass CAPTCHAs is rare enough it seems reasonable to ask those people to use a browser plugin.