This is more about how well (or not) climate models predict/project the total warming that's been seen so far.
After I put up this graph of a GISS model's cumulative warming versus observations, a couple of people said "but it's just one model." Fair enough. It happened to be the first model I looked at, because I'd been reading a paper about it.
The same CMIP5 page also has a data file for an ensemble of 21 model -- presumably, the "best" ones, the biggest ones, made by the big research groups. (All models aren't created equal, which makes one wonder about this Roy Spencer graph with an average of 90 models. More about that graph at HotWhopper.)
Here's the result for the total warming from 1860 to 2014:
C&W is Cowtan and Way. Again, by "total warming" I mean (linear slope)*(time interval). (Not perfect, but what would be better? Maybe average_of_last_30_years - average_of_first_30_years. Maybe I'll try that later.)
So, the agreement is pretty good. The deviation between the ensemble mean and Cowtan & Way starts around 2005, and is now up to 0.07°C. There's your slowdown.
If I was Stephen Koonin, I would say the model error is only 0.07 K/288 K = 0.02%. But I'm not, so I won't.
Instead, it's about 0.07 K/0.94 K = 7%, over a century and a half.
So how good does a model need to be?