Fractured carbonate reservoirs are typically modeled in a system of dual-porosity and dual-permeability (DP/DP), where fractures, vugs, karsts and rock matrix are represented in different domains. The DP/DP modeling allows for a more accurate reservoir description but implies a higher computational cost than the single-porosity and single-permeability (SP/SP) approach. The time may be a limitation for cases that require many simulations, such as production optimization under uncertainty. This computational cost is more challenging when we couple DPDP models with compositional fluid models, such as in the case of fractured light-oil reservoirs where the production strategy accounts for water-alternating-gas (WAG) injection. In this context, low fidelity models (LFM) can be an interesting alternative for initial studies. This work shows the potential of compositional single-porosity and single-permeability models based on pseudo-properties (SP/SP-P) as LFM applied to a fractured benchmark carbonate reservoir, subject to WAG-CO2 injection and gas recycle. Two workflows are proposed to assist the construction of SP-P models for studies based on (i) nominal approach and (ii) probabilistic approach of reservoir properties. Both workflows begin with a parametrization step, in which the pseudo-properties are optimized for a base case in order to minimize the mismatch between forecasts of the SP/SP-P and DP/DP models. The new parametrization methods proposed in this work showed to be viable for the construction of the SP/SP-P models. For studies under uncertainties, the workflow proposes obtaining pseudo-properties by robust optimizations based on representative models from a DP/DP ensemble, which proved to be an effective method. The case study is the benchmark UNISIM-II-D-CO with an ensemble of 197 DP/DP models and two different production strategies. The risk curves for production, injection and economic indicators obtained from DP/DP and SP/SP-P ensembles showed good match and the computational time spent on simulations of the SP/SP-P ensemble was 81% faster than DP/DP models, on average. Finally, the responses obtained from both ensembles were validated in a reference model (UNISIM-II-R) that represents the true response and is not part of the ensemble. The results indicate the SP/SP-P modeling as a good LFM for preliminary assessments of highly time-consuming studies. Besides, the workflows proposed in this work can be very useful for assisting the construction of SP/SP-P models for different case studies. However, we recommend the use of the high-fidelity models to support the final decision.