Traynor was quoted in a networkworld article last year saying they aim for three and a half nines (99.95%). But you need to read into the incidents more carefully -- figuring out actual "uptime" is quite hard. Consider the longest-lasting incident:

  "On Tuesday 23 February 2016, for a duration of 
   10 hours and 6 minutes, 7.8% of Google Compute Engine
   projects had reduced quotas.  ...  Any resources that
   were already created were unaffected by this issue."

I'm not sure off the top of my head how I'd try to compute the overall availability #s from that one. One can possibly try to determine and sum the effects on the individual customers, but we can't from the information provided. But it's certainly less overall downtime than just counting it as a 7 hour failure.