Rock-Eval® (RE) is a thermal analysis technique increasingly used to characterise soil organic matter. To interpret the results, particularly when investigating differences between samples, it is necessary to know the expected ranges of analytical error associated with the RE measurements. Moreover, the RE analyzer is now at its seventh version (RE7) while most literature results were produced using the previous version (RE6). Thus, a characterization of the reproducibility of RE measurements is necessary. We measured the reproducibility of RE measurements using fifteen samples from French croplands and forests that were analysed on five different RE instruments, located in different laboratories and belonging to both generations RE6 and RE7. From each RE analysis, we extracted RE parameters commonly used for soil organic matter characterization and we performed the prediction of the active and stable soil organic carbon fractions using a machine learning model (PartySOC) that uses RE parameters. We obtained a measure of the expected relative errors in RE parameters and PartySOC predictions per instrument, across instruments of the same generation and across generations. We found that the parameters total organic carbon (TOC), mineral carbon (MinC) and R-index are well reproducible, even across the RE6 and RE7 generations. Instead, the hydrogen index (HI) and oxygen index (OI) are more sensitive to signal variations, even within the same generation, especially when TOC is low. The PartySOC predictions were well reproducible across RE6 instruments but not across RE generations. In the future, the results of this study will help discriminate relevant differences between soil samples characterised using RE thermal analysis.