Classical algorithms used for travel time tomography are not necessarily well-suited for handling very large seismic data sets, densely parameterized velocity models and for taking advantage of current supercomputers architecture. We revisit the classical approach of first-arrival travel time tomography by proposing to use a simple gradient-based approach without the need of estimating the inverse of the Hessian of the misfit function. The gradient-based approach is very attractive from a numerical point as it can handle large data sets and velocity models and is very robust with respect to the choice of the initial velocity model.