I am comparing two models, I have only one accuracy per model. So I wonder, is it correct to generate for example 10.000 bootstrap samples (accuracies) and use the bootstrap samples to make a t-test for comparing the models?
No. Tests for hypotheses of mean difference using the bootstrap is covered elsewhere. For instance, here: Computing p-value using bootstrap with R
You can't confuse the bootstrap resample as being an observation. This is because the number of bootstrap re-samples is arbitrary. If the hypothesis you wish to test is that the mean is statistically significantly different between the two groups, you can increase the number of bootstrap samples so that any non-zero difference will render a statistically significant result.