Shall I trust AIC (non-full model) or slope (full model)?

The purpose to run regressions for butterfly richness again 5 environmental variables is to show the importance rank of the independent variables mainly by AIC.

In non-full models, they reveal that variable A tends to be more influential than the others by delta AIC.

However, in the full model, the regression coefficient of variable A is slightly second to that of variable B. (R-square of the full model is 1.43)

The conflicting outcomes (non-full model by AIC and full model by slope) seems to make it difficult to ascertain that variable A is the variable mostly weighted.

Please kindly suggest which criterion should be relied on for the specified purpose or any further test should be carried out. Thank you.

R-Squared can't be 1.43... and other errors make your question hard to interpret.

Here's a sort of generic response that might eventually lead to an answer.

The AIC score tells you how good the model is similar to R-squared but penalizes it based on how many components are in the model. You can theoretically always get a better fit with more elements to the model, and R-squared reflects that, but at some point adding more explanatory variables doesn't increase the model accuracy as much as it needlessly increases complexity. Therefore, if you add a factor and the AIC goes down instead of up, that's because you're not adding enough explanatory power to make up for the increased complexity of the model and you should err on the side of parsimony (i.e. remove that factor).