Thank you for the manuscript “The North Atlantic mean state in eddy-resolving coupled models: a multimodel study”. Understanding how increasing resolution improves the mean state is an important topic and I welcome this study. However, I have some major comments so have to recommend major revisions.

There are a lot of statements throughout about quantities being larger/smaller in the HR ensemble compared to the HR ensemble, however this isn’t really tested. It seems to be based on differences in the ensemble mean without taking into account the fact that the two ensembles are quite different in size and possibly quality. If 4 members were picked at random from the LR ensemble, what is the probability that they look like the HR ensemble? It should also be borne in mind that the CMIP6 ensemble includes some models which are a long way from the observations. The authors should statistically test their assertions – one way of doing this is by using single number metrics (for instance the AMOC strength, SST in the CNA etc) and testing whether the ensembles are statistically different by: randomly picking 4 members from the LR ensemble, calculating the ensemble mean, then repeating until you have a distribution from the LR ensemble. This will show whether the HR ensemble is really different from the LR ensemble, or whether there are some LR members which have similar properties.

My other concern is that the analysis here is quite basic and doesn’t really show much that is new. The authors could include more analysis of scatterplots of the metrics they analyse against each other and discuss the implications for how biases affect each other. There have been a number of studies looking at how resolution affects the North Atlantic. The novelty of this paper seems to be having multiple models at eddy-resolving rather than permitting resolution. What are the implications of going to eddy-resolving resolution?

Minor

L11 ‘important role in featuring global ocean dynamics’ – what does this mean?

L19 ‘weaker than for lower resolution models’ This is rather misleading – the way this is reported in the abstract suggests that it is a result of the resolution change. As the authors discuss, studies have shown that the impact of increased resolution varies from model to model. The result here that the AMOC is weaker in the LR ensemble is likely because of some very strong models in the LR ensemble.  

L63 ‘to a’-> ‘with a’

L105 Include resolution in km to compare with other values.

L173 I can’t see the paper ‘in review’ though other studies (eg Marzocchi et al 2015) have suggested that this bias is because of the location of the NAC , rather than the strength.

L237 and Fig 6 Using correlation of profiles to assess the shape seems rather flawed – if one profile has twice the slope of a second profile then the correlation would be perfect. It would be better to assess the stratification itself.

L243 ‘expected to impact on’

Fig 6 What region are these for?

L256 ‘Its’ -> ‘The’

Fig 8 Might be clearer if you adjusted the scale

L287 I don’t really understand this sentence.

L305 It’s not entirely clear what you mean by methodological approach and what it affects – expand a little on this.

L331 ‘biases in NCH’

L347 ‘reduced cold and fresh biases’

L380-383 and Fig A1 Needs more discussion. The correlations in Fig 1a are certainly not significant so I’m not sure what it means that the correlation increases. Maybe instead discuss why there might be correlations and why the HR models might be outliers.

L384-386 Doesn’t really make sense – what is ‘it’?

L397 temperature signals are often damped by interactions with the atmosphere reducing the amplitude of the signal. This would explain why there are stronger correlations for salinity

L401-404 I don’t understand what the authors are getting at here – please explain more.

L410-418 Different definitions of MLD can give very different results. Do all the studies mentioned use the same density criteria? Also you don’t make it very clear that using monthly mean densities to calculate a MLD can give very different (and likely shallower) estimates than means of instantaneous profiles used by Yashayaev etc.

L463 ‘southward propagation of the MLD signal into the AMOC’ – what does this mean?