I guess it's kind of clear from their websites whether they're actually building models vs using something off the shelf. OpenAI and Deepmind both regularly publish papers and technical blog posts. If you don't see that on the company's website that's one indication.

I'd just check: are they publishing technical stuff you barely understand? Green flag. Is the general idea exciting? Another green flag.

It's a faulty heuristic because more than a few good ones will fall through the cracks but oh well