This is a special post for quick takes by J Bostock. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I recently ran a number of autoresearch runs with DeepSeek V4 (I wanted the reasoning traces and was considering doing some RL to make a hacking model). When using DeepSeek themselves, I noticed that ~10% of my responses weren't coming back with thinking, even though thinking tokens were being billed. The obvious answer here is that it was Claude.
I trust Tinker not to do this. I don't really trust anyone else. So I suppose it's Tinker for me from here on out for API access on open-weights.
EDIT: I also trust the other "big" companies, AWS and Google Cloud-type stuff, when they serve models. Those are often somewhat behind the frontier, though.