Ars Technica has this quite accessible post on deepfakes.
Conclusion: not there yet. But probably not too far off either.
Related: Tom Scott's video on the same subject, coming to the same conclusion, except with much weaker hardware and hence much less training.
The Ars Technica promoted comment also makes a good point: entities with much more massive budgets than $500 over two weeks, and hence more powerful hardware, could probably use the technology even as it is now, perhaps with some tweaks, to perform faceswapping at a much more convincing level. The scale available to them is unimaginable to us, and nearly nothing prevents them from using it for whatever purpose they might want.