Ecosystem models forced by future climate simulations outputs from the Coupled Model Intercomparison Project Phase 5 (CMIP5) simulate a substantial decline of tropical marine animal biomass over the course of the 21st century. Regional projections are however far more uncertain because of well-known biases common to most CMIP5 historical simulations that propagate within the food web. Moreover, the model outputs for high trophic levels marine fauna suffer from lack of validation based on in situ data. In this study, we implement a "bias-mitigation" strategy to reduce the physical oceanography and biogeochemical biases simulated by three CMIP5 models under the future RCP8.5 scenario. We force two very different micronekton models with these "bias-mitigated" outputs to infer the future micronekton changes in the Coral Sea: a 3-D deterministic population dynamics model; and a 3-D statistical model based on in situ hydro-acoustic data. These two models forecast a consistent pattern of micronekton abundance changes in the epipelagic layer (0-150 m) by 2100 for three different climate forcing used, with a marked decrease south of 22 degrees S and a smaller increase further north mostly related to temperature and chlorophyll changes. In contrast, changes in the vertical patterns of micronekton predicted by the two models considerably differ in the upper mesopelagic layers (150-450 m) and lower mesopelagic layer (450-1000 m), highlighting the structural sensitivity in model type. Since micronekton are prey of all larger marine predators, those discrepancies in vertical structures of micronekton may hamper our potential to predict how top predators may evolve in the future.