The exposure-lag response of air temperature on COVID-19 incidence is unclear and there have been concerns regarding the robustness of previous studies. Here we present an analysis of high spatial and temporal resolution using the distributed lag non-linear modelling (DLNM) framework. We first fit statistical models to select Italian cities, accounting for lag effects up to 10 days and several categories of potential confounders (policy, mobility, meteorological, and pollution variables). Estimates from these models are then synthesised using random effects meta-analysis to yield pooled estimates of the exposure-lag response presented as the relative risk (RR) and cumulative RR (RRcum). Though there was variation in the lag-specific exposure-response curves, the cumulative exposure response was approximately bell shaped, with highest risk at 19.8 °C, 2.39 [95% CI: 1.13; 2.94] times that at 4.7 °C which represented the lowest risk. Our work is in agreement with studies suggesting “lower” and “higher” temperatures might reduce covid-19 transmission, though our results suggest the optimum temperature for outdoor transmission might be higher than previously thought. Due to this uncertainty, our work underscores the need for facemasks and social distancing even in warm temperatures.