We propose a new framework allowing to estimate non-parametric affine and non-affine option pricing models. Our method applies autoencoder neural networks to the log-characteristic function implied by observed option prices. Since the logarithm of the characteristic function is linear in the state factors under the affine assumption, we obtain a data-driven affine model by specifying a linear mapping in the autoencoder architecture. Alternatively, we let the data speak about any needed non-linearities to estimate a non-affine model. Using an extensive panel of S&P 500 option data, our approach reveals that the non-affine class of models significantly outperforms the affine class in pricing options out-of-sample.