The autocorrelation function is a bit odd because the autocorrelations do not subside as they should do. That suggests that there is no adequate AR model. The partial autocorrelation function is more helpful. The highest peaks seem to be a lags 2 and 3. So I would start with those lags. Or you could do something else.
Determining the model order from the autocorrelation and partial autocorrelation is a bit of an art. I believe that it is not necessary to do that these days. Computers are so fast that it is no problem to estimate severals models and pick the best one. In my opinion that is the way how to determine the model order these days. Using the autocorrealation and partial autocorrrelation functions seems to be the correct method of the old days when estimating a model was a lot of work.
So, estimate several models using a loop and calculate the Akaike information criterion for each model. Then pick the model with the lowest AIC. The process is also described in this Matlab example. Don't forget to check the residuals. They must be free of autocorrelation.
